Additive manufacturing (AM) has spent much of its history inside controlled environments. Research labs, prototyping departments, and specialist teams could tolerate complexity because expertise was concentrated and production expectations were limited. Today, that context has changed. Across industries from aerospace to defense, industrial equipment, and spare parts, AM is increasingly expected to function as a production technology rather than an experimental capability.This transition has exposed a structural imbalance. 3D printer performance and material portfolios have matured steadily, yet the surrounding software infrastructure has struggled to keep pace. While specialized software has been good enough for small groups of highly skilled specialists, it now has to support distributed manufacturing organizations, integrate with enterprise systems, and operate at scale without constant expert intervention. In many deployments, software has emerged as the primary bottleneck for growth.These observations frame how Materialise is approaching the next phase of AM. In a conversation with Bart Van der Schueren, Chief Strategy and Technology Officer, and Tom Craeghs, Director of Product Management, they shared that scaling AM requires the rethinking of software architecture itself, grounded in how manufacturing organizations actually operate today.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1770076878594-Screenshot+2026-02-03+000035.png" style="width:100%px" class="fr-fic fr-dib" alt="Bart Van der Schueren (left) and Tom Craeghs (right) from Materialise" />Bart Van der Schueren (left) and Tom Craeghs (right) from Materialise<h2>From Isolated AM Departments to Distributed Production</h2>For much of its early evolution, AM software was designed as a technical toolbox. These environments assumed deep user knowledge and allowed engineers to experiment directly with parameters, geometry preparation, and process tuning. That model functioned well in protected settings, where a small number of specialists controlled the entire workflow.AM is now moving into mainstream production contexts. New roles are involved, including operators, planners, quality engineers, and sales teams. These users interact with AM as part of a broader digital manufacturing thread rather than as an isolated process. Software that depends on continuous expert intervention becomes difficult to scale under those conditions.The challenge is not simplification in the sense of eliminating complexity from the process. AM remains inherently complex. Instead, the challenge lies in presenting that complexity in a way that different roles can interact with reliably, without requiring everyone to become an AM specialist. This shift changes what software has to do. It must orchestrate workflows, preserve expertise, present a sufficient level of explainability, and integrate seamlessly with other manufacturing systems.These requirements represent a structural integration issue, especially when deploying AM at scale. AM software has historically developed its own domain language, shaped by build preparation, slicing, and process parameters. Manufacturing IT systems operate with varying abstractions, focusing on orders, routings, quality records, and traceability.When these two worlds fail to align, integration becomes fragile. Data moves between systems through manual steps or brittle interfaces, limiting automation and increasing operational risk. Bridging this gap requires going beyond technical connectors and into a shared architectural model that allows AM concepts to be expressed and understood within the broader manufacturing context.Materialise treats this mapping as a first-order design problem. The goal is to allow AM to function as a native component of manufacturing operations, rather than as a specialized exception that sits alongside them.<h2>Rethinking the Role of End-to-End Platforms</h2>Industry discussions around AM software have often emphasized end-to-end platforms. The promise has been consistency and control across the workflow, from design intake to finished part. Materialise’s experience with industrial deployments has led to a more nuanced position.Van der Schueren describes this approach using a T-profile analogy. The horizontal axis represents an end-to-end workflow that can take a design through production with full traceability. The vertical axis represents deep AM expertise, built from decades of algorithm development and process knowledge.Some organizations benefit from adopting the full horizontal workflow, particularly when they lack existing manufacturing systems. Others already operate mature ERP, MES, and PLM infrastructures and only require access to additive-specific capabilities. Materialise’s software architecture is designed to support both cases, allowing companies to adopt a complete workflow or integrate selected components into their existing systems.This architectural choice reflects an acknowledgment of manufacturing reality. Most industrial environments are heterogeneous. For AM to scale, software must accommodate that diversity rather than attempt to replace it.At the foundation of this approach sits Materialise Magics, which continues to serve as the production-proven core for data and build preparation. Building on this foundation, the Co-AM platform extends Magics into an open, modular software architecture that connects workflows, users, machines, and enterprise systems. Within this platform, Brix operates as an automation and orchestration layer, enabling customers to translate deep AM expertise into executable, reusable workflows rather than fixed sequences of manual steps.Materialise summarizes this philosophy internally as “your process, your rules.” It reflects a deliberate decision to let customers define how AM fits into their operations, instead of forcing diverse production realities into predefined software workflows.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1770077650138-hm-materialise-magics.jpg" style="width:100%px" class="fr-fic fr-dib" alt="Materialise Magics, Software for 3D Printing Data and Manufacturing Preparation (Credit: Materialise)" />Materialise Magics, Software for 3D Printing Data and Manufacturing Preparation (Credit: Materialise)<h2>Why Additive Manufacturing Resists Abstraction</h2>One reason AM software has proven difficult to generalize lies in the physics of the process itself. Unlike conventional manufacturing, where material properties are largely defined before shaping occurs, AM produces geometry and mechanical performance simultaneously. Orientation, support strategy, scan parameters, and thermal history interact continuously, shaping both the form of a part and its functional behavior.This coupling creates a level of process interdependence that resists simplification through generic abstractions. Van der Schueren notes that this reality fundamentally shapes how Materialise approaches usability. The objective is not to eliminate complexity but to make it legible and controllable. Software must expose the right degrees of freedom without requiring every user to engage at the same technical depth.These constraints manifest differently across production contexts. In high-mix, low-volume applications such as prototyping, on-demand spare parts, and certain defense programs, the central challenge tends to be decision reliability. Organizations must assess feasibility consistently, prepare builds predictably, and reuse validated process knowledge across a large number of unique parts. Scalability in these environments depends less on throughput than on how effectively expertise can be formalized and distributed beyond a small group of specialists.Series production introduces a distinct economic pressure. While AM can support production volumes in the thousands and, in some cases, even millions of parts, most applications operate well below the scale at which traditional manufacturing amortizes new product introduction (NPI) costs. Even when volumes reach the tens of thousands, NPI effort often remains a dominant factor in the business case. This explains why many AM applications succeed technically yet struggle to remain economically viable once sustained production begins.From Materialise’s perspective, this is where software becomes a decisive lever. Many business cases do not fail because AM cannot deliver functional parts but because the path from development to economical production remains too long or too opaque. Software that reduces the effort required to optimize parameters, refine support strategies, and iterate on design and process choices can shorten the NPI phase and make incremental cost improvements meaningful at moderate volumes. In this context, higher-volume AM becomes less about reproducing mass production economics and more about systematically lowering the threshold at which series production becomes economically viable.Across both low-volume and series contexts, customers increasingly seek ownership of their processes. Moving beyond fixed parameter sets allows organizations to adapt AM to their specific technical and economic constraints, rather than conforming their applications to predefined workflows.<h2>Co-AM Brix and the Operationalization of Expertise</h2>Materialise’s investment in Brix addresses this demand for process ownership directly. Brix is a visual programming environment built on Materialise’s software development kit (SDK), which now contains more than a thousand AM algorithms developed over more than three decades.Rather than exposing these capabilities exclusively through expert-centric tools, Brix allows process engineers to assemble validated workflows visually, observe outcomes immediately, and deploy those workflows across an organization. In high-mix, low-volume environments, this enables organizations to automate decision-making and standardize preparation steps without diluting expertise.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1770077678168-1770077678168.png" width="602" height="339" class="fr-fic fr-dii" alt="CO-AM Platform from Materialise (Credit: Materialise)" />CO-AM Platform from Materialise (Credit: Materialise)A process engineer can, for example, encode feasibility checks, orientation logic, or quality constraints into a Brix workflow. Sales teams or front-line engineers can then apply that workflow automatically to incoming parts, receiving consistent assessments without requiring direct expert intervention. Standards are preserved, response times improve, and expertise scales across the organization in a controlled manner.In series production contexts, Brix serves a complementary function. Engineers use it to explore interactions between design choices, support strategies, and process parameters within a unified environment. While the underlying complexity remains, its behavior becomes observable, allowing development teams to understand trade-offs more clearly and converge on robust production strategies faster.This capability is particularly relevant for original equipment manufacturers (OEMs) that already operate internal AM centers. As Craeghs explains, these groups often serve as incubators for new applications, supporting prototyping and limited production long before series manufacturing is viable. Materialise’s role is to help these organizations translate internal expertise into repeatable, scalable workflows that can survive the transition from experimental capability to sustained production. By reducing operational friction and uncertainty, software becomes the mechanism through which AM moves from isolated competence to industrial infrastructure.<h2>A Gradual Transition Shaped by Product Redesign</h2>Van der Schueren and Craeghs describe the current phase of AM as a measured transition rather than a sudden inflection. The most successful applications emerge when products are redesigned specifically for AM, not adapted from conventional designs. Such redesign cycles take time.As products developed over the past decade begin to move into production, their success creates a multiplier effect. One validated application leads to additional machines, more automation, and deeper integration into manufacturing operations. That progression reinforces the need for software architectures that can scale without centralizing expertise.Materialise positions itself within this evolution as a solution provider. Its focus lies in translating deep AM knowledge into systems that organizations can operate, adapt, and trust over time.The long-term trajectory of AM depends on the accumulation of reliable, repeatable production experience. Software sits at the center of that accumulation, defining how knowledge is captured, workflows executed, and AM connected to the rest of the factory. This evolution requires openness, modularity, and an architectural discipline informed by decades of domain expertise. With AM’s gradual integration into production, the companies that succeed will be those that treat software as the foundation upon which scale is built.For more information about Materialise’s approach to AM, visit <a href="https://www.materialise.com/" target="_blank">www.materialise.com</a>

How Materialise is redefining software architecture as additive manufacturing moves into industrial production

Re-Architecting Additive Manufacturing for Scale

A new study uses advanced imaging technology to identify materials in municipal solid waste that can be separated for recycling or to produce energy. The study made use of hyperspectral imaging, a method that uses powerful optical sensors which capture the light spectrum across every pixel in an image. By analyzing the ways that different materials reflect light even outside of the visible spectrum, hyperspectral imaging enables researchers to create unique spectral “fingerprints” for each individual material, allowing for fast identification of materials that might look identical to the naked eye. “Hyperspectral imaging is a powerful tool that allows us to see what human eyes or standard cameras can’t,” said Lokendra Pal, E.J. Woody Rice Professor and University Faculty Scholar in the Department of Forest Biomaterials at North Carolina State University and a co-author of the study. “With this technology, we can capture real-time images of large quantities of waste, down to the pixel level of data. By doing that, we can identify different materials based on variations in light reflection that we could not normally see.”Hyperspectral imaging also allows scientists to determine not only the material type, but how much of it there is and whether it is contaminated, Pal said. This helps make recycling operations more cost-effective and efficient.Humans see light on what is known as the RGB spectrum, standing for red, green and blue. Light within this spectrum has wavelengths of roughly 400-700 nanometers, which our eyes perceive as color. Hyperspectral imaging is able to capture wavelengths up to 2,500 nanometers, covering the near-infrared and shortwave infrared ranges. This creates a tremendous amount of data, which can be leveraged with machine-learning to identify waste materials that can be converted into valuable products. “For example, coffee cups are made from plastic and paper,” said lead author Mariangeles Salas, a Ph.D. student in the Department of Forest Biomaterials at NC State. “Millions of these cups are thrown away each year with less than 1% recycled. “With hyperspectral imaging, we create what is known as a data cube,” Salas explained. “This is a visual representation which describes a pixel’s unique light reflection characteristics in three dimensions. This allows us to identify subtle differences between materials, such as two types of paper in the same coffee cup. Both contain cellulose, but their chemistry and composition differ, meaning they are better suited for different recycling pathways.”<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1759149183513-HSE-Data-Cube_2-1-1536x788.png" style="width:100%px" class="fr-fic fr-dib" />A data cube created with hyperspectral imaging.Researchers intend to put this huge influx of data to broader use by creating one of the largest libraries of visual and hyperspectral images with detailed metadata of municipal solid-waste materials. With over a billion spectral pixels and counting, this open-access repository of data will provide waste managers such as municipalities, materials recovery facilities and researchers with an invaluable tool. This technology could help speed up and improve the accuracy of automated recycling systems, increasing efficiency and reducing the amount of recyclable material lost to landfills, and support a more sustainable circular economy. The study, <a href="https://www.sciencedirect.com/science/article/abs/pii/S2590238525004084#preview-section-references">“Hyperspectral imaging for real-time waste materials characterization and recovery using endmember extraction and abundance detection,”</a> is published in Matter. Co-authors include Mariangeles Salas, Simran Singh, Raman Rao, Raghul Thiyagarajan and Lucian Lucia of NC State University; Ashutosh Mittal and John Yarborough of the National Renewable Energy Laboratory; and Anand Singh of IBM. This work was supported by the U.S. Department of Energy’s Office of Energy Efficiency and Renewable Energy (EERE) under the Bioenergy Technologies Office (BETO) award number DE-EE0009669.

A coffee cup being photographed with a hyperspectral camera. Hyperspectral imaging can help scientists locate subtle variations in materials within large amounts of solid waste.

A new study uses advanced imaging technology to identify materials in municipal solid waste that can be separated for recycling or to produce energy.

How Hyper-detailed Cameras Will Make Recycling More Efficient

<h2 id="isPasted">Introduction</h2>The ability to combine both pictures and words is a powerful capability in many of the latest deep learning models. These models can analyze the pixels of your image to understand its content, analyze the words of your prompt to grasp what you are asking, and then fuse those understandings together to generate a response. While these models still have their limits, their capabilities would have been unimaginable only a few years ago.CLIP, a text-image model published by OpenAI, has been the most impactful approach to fusing text and images. It is a useful standalone model that also underpins the vision-language capabilities in many of today’s multi-modal models, from Large Language Vision models like <a href="https://llava-vl.github.io/">Llava</a> to generative models like <a href="https://en.wikipedia.org/wiki/Stable_Diffusion">Stable Diffusion</a>. Here, I’ll explore how CLIP works, its practical applications, and its limitations.<h2>The Use-Case</h2>Traditional image classifiers pre-define a fixed set of image categories and are trained using examples to classify images into those categories. It’s an effective process if you just want to distinguish between “cat” and “dog” pictures – but what if you need a model that works for arbitrary labels? Instead of fixed categories, could a model be trained to understand arbitrary text descriptions of categories?<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1757423180979-imageclassifier.webp" style="width:100%px" class="fr-fic fr-dib" /><img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1757423196650-imageclassifier2.webp" style="width:100%px" class="fr-fic fr-dib" />A pre-trained classifier might have labels for different animal types, but it won’t know what sounds animals make or which one spins a web.Before multi-modal text-vision models like CLIP, these fixed-label image classifiers were the best we had. Datasets with millions of images were manually tagged with a huge fixed set of labels like “cat”, “spider”, “fluffy”, etc., and trained with supervised learning techniques. This method works for those pre-defined labels but isn’t scalable. There are too many possible categories. These models don’t effectively learn the relationships between words or answer queries that don’t match their training data. They also don’t leverage the similarity between words to efficiently generalize their understanding.At the same time, research in both natural language and computer vision already had robust concepts of embeddings – representing words, sentences, and images as vectors, a sequence of numbers that encode “relevant” features such that similar objects have similar encodings. Embeddings for “cat” and “dog” would be closer to each other than to the embedding for “skyscraper” – whether they are embeddings for the words themselves or images portraying them. These embeddings give us a way to measure the ‘semantic similarity’ of two things by looking at the distance between their embeddings. However, embeddings are only comparable to other embeddings generated by the same model. The embedding of a cat image using an ImageNet derived model couldn’t be meaningfully compared to the word2vec embedding of the word “cat.” <img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1757423241446-1757423241446.png" class="fr-fic fr-dib" />Similar concepts are nearby in each embedding space, but you cannot compare different ones.<h2 id="isPasted">CLIP – Contrastive Language-Image Pre-training</h2>In 2021, OpenAI released their <a href="https://github.com/openai/CLIP">“Constrastive Language-Image Pre-training” (CLIP)</a> model. The basic idea is to take an internet-scale dataset of captioned images and train a model to generate the same embedding for the caption as it does for the corresponding image. The result is a model that embeds both text and images into a shared feature space.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1757423277719-Combining-Language-and-Vision_Final-2.webp" style="width:100%px" class="fr-fic fr-dib" />The caption “cat” and the corresponding picture map to the same embedding.This immediately enables powerful applications. For example, we can generate an embedding for an image, compare it to a list of text embeddings, and determine which text best describes the image. This approach allows for classification based on arbitrary phrases instead of a predefined set of categories.Getting more technical for a moment, the CLIP training process uses contrastive loss to ensure that the embeddings of a caption and its corresponding image end up near each other and far away from other unrelated captions and images.  For each training step, the model input is a batch of N (image, caption) pairs, and the output is an embedding of each image and caption. The training objective is to maximize the cosine similarity of the true image-text pairings and minimize the similarity of all other pairings in the batch.  This training process is the important part – the actual model architecture of the text encoder and image encoder is an interchangeable implementation detail.This simple process works remarkably well. Much of the <a href="https://arxiv.org/abs/2103.00020">CLIP paper </a>describes the new state-of-the-art results it can achieve across a variety of computer vision tasks and applications. It isn’t just memorizing image-caption pairs during training, but instead developing a shared representation of image and text that generalizes beyond the examples it was trained on. Points in this feature space encode both an image and the words describing that image, which has very powerful applications.<h2>CLIP as a Classifier – Applications and Limitations</h2>For images and texts similar to the ones it was trained on, CLIP works almost like magic.  If I want to distinguish between “red cars” vs “blue cars” or “sports car” vs “SUV,” the model just works. Where previously I needed to collect examples and train an application-specific image classifier, now I can just describe that I want to distinguish between an “open door” and a “closed door,” or tell me, “is something on fire in this image?”, and I instantly have something usable.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1757423300089-doors.webp" style="width:100%px" class="fr-fic fr-dib" />Glad I didn’t need to go collect a dataset of closed door and open door examples!For more complex, domain-specific detections, off-the-shelf CLIP models are not a standalone solution. They are still limited by their training data and by their pretraining objectives. Remember that these models are trained on “captioned images on the internet.” The more dissimilar your domain is from that, the worse these models will perform. Be aware that many examples you see online about this model (including in this article!) are cherry-picked to make their point.If you try to use one of these models to detect subtle anomalies in your manufacturing process, you will be disappointed. It will not tell you the physical dimensions of your dent, and may struggle to differentiate between a scratch and the normal texture of your material. CLIP does not train well for spatial reasoning.  They can often distinguish between an image containing “a couple of cats” vs one with “tons of cats,” but don’t expect them to accurately count exactly how many cats are in the image.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1757423319429-cat.webp" style="width:100%px" class="fr-fic fr-dib" />CLIP doesn’t know how tall the cat in this picture is.Resolution is also an important limiting factor. CLIP has some limited OCR capabilities, but it is hindered by the low resolution it operates at – often as low as 224×224 pixels – preventing small text from being legible. Other fine details in images are lost during resizing, making them inaccessible to the model when applied naively.<h2>Thresholding and Relevance</h2>Another challenge with using CLIP and similar models is interpreting the ‘similarity scores’ it outputs. Anyone who has worked with detection models is familiar with confidence scores. While these are often uncalibrated and mistakenly interpreted as probabilities, they are at least consistent enough that you can use rules of thumb or develop an intuition as to what constitutes a ‘likely’ detection. The same is not true for CLIP models, particularly as your queries deviate from what the model was trained on.Many demonstrations of this technology look at relative scores. Is the “cat” score higher than the “dog” score in this image? Or which image in my dataset has the highest “zebra” score? Looking at relative scores instead of absolute can avoid this issue, but is not always a useful solution.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1757423342072-cat2.webp" style="width:100%px" class="fr-fic fr-dib" />This cat is very #FF00FF’y. But you wouldn’t be wrong if you called it violet or purple.To robustly set similarity score thresholds, you really want to put together a dataset specific to your use-case and generate standard evaluation metrics – but then we’re back to labeling and curating datasets, which we wanted to avoid.Finding clever and general ways to solve this problem is key to making use of CLIP directly. Fortunately, many downstream tasks don’t need to solve this problem to make use of CLIP embeddings and take advantage of their fused image/text representations.<h2>CLIP for Semantic Similarity</h2>There are many systems out there using CLIP as the basis for their text-image similarity searches. These embeddings are a great solution for efficiently filtering images from a massive collection based on text descriptions. If I want to find that “night-time picture of a cat sleeping on a snowman” that I know is in my collection somewhere, it works like a charm!<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1757423372042-snowman.webp" style="width:100%px" class="fr-fic fr-dib" />I knew I had that picture somewhere!These systems work by using CLIP’s image encoder to generate embeddings for every image in the collection and inserting those embeddings into a vector database. Once that is done, any text query can be very quickly answered by generating the text embedding and searching our database for the N most similar image embeddings. Vector databases are incredibly efficient at this operation.Of course, most of the same limitations of using CLIP as a classifier apply here, too. Thresholding is still important for determining where to cut off the results. The effectiveness of the search will often depend on how unique your dataset is and how complex the queries you need to match. Subtleties are not captured by the model, and the metric often struggles to correctly order results when many of them are close. But even with those caveats, it’s still remarkably good and wildly better than what came before it.<h2>Conclusion</h2>Since this groundbreaking approach was first published, the research community has made many improvements. While the core concepts remain the same, researchers have taken these ideas and refined them, making models that are more accurate, faster, or in some other way better. Additionally, many diverse and unexpected applications of the model have been found. CLIP-extracted features are used as the basis for video understanding, prompt-guided image generation, large language vision models, zero-shot segmentation models, and more.While we don’t have time here to get into all of those extensions to the model, we’ve covered the basics here to understand what you can do with CLIP and how it works. While it cannot reliably solve everything by itself, it is a valuable tool in your computer vision toolbox and a great starting point for more powerful capabilities.At Matroid, part of what we do is keep on top of these advances in AI and computer vision and figure out how the ideas and concepts can be applied to solve the industry-specific use-cases encountered by our customers. Combining language and vision with techniques like those pioneered by CLIP enables a lot of capabilities – we work to transform these innovations into reliable, practical solutions powered by our computer vision platform.

CLIP, a text-image model published by OpenAI, has been the most impactful approach to fusing text and images.

Combining Language and Vision – Looking at CLIP

A new blue fluorescent molecule set new top emission efficiencies in both solid and liquid states, according to a University of Michigan-led study that could pave the way for applications in technology and medicine.&nbsp;Able to absorb light and emit it at lower energy levels, fluorescent molecules called fluorophores glow in OLED displays and help doctors and scientists figure out what’s happening in cells and tissues. They need to be solid in displays and many sensing applications, but liquids are typically preferred for biological uses. Most fluorophores don’t work well in both forms, but this one does.“The fluorescent material reached record-breaking brightness and efficiency with 98% quantum efficiency in the solid state and 94% in solution,” said <a href="https://mse.engin.umich.edu/people/jinsang" target="_blank" rel="noreferrer noopener">Jinsang Kim</a>, the Raoul Kopelman Collegiate Professor of Science and Engineering in the U-M Department of Materials Science and Engineering who led the study, which is published in Nature Communications.Often, engineers designing fluorophores start in solution, exploring the optical properties of individual molecules, but run into problems in their solid-state applications when fluorophore molecules contact each other.&nbsp;“Fluorophores behave very differently in the solid state, which then requires more rational molecular engineering effort for structural modification,” Kim said “By investigating and establishing a molecular design principle to make fluorophores that are bright both in solution and solid states, we have reduced development time and cost for various future applications.”The initial discovery of the versatile fluorophore—called TGlu for short—was unexpected for lead author <a href="https://www.linkedin.com/in/jungmoo-heo-4479351bb/" target="_blank" rel="noreferrer noopener">Jung-Moo Heo</a>, U-M postdoctoral research fellow of materials science and engineering.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1751417438256-TGlu_Liquid-Solid-featured.v1.jpg" style="width: 100%px;" class="fr-fic fr-dib">The new TGlu molecule is a small but powerful single benzene fluorophore with a quantum yield—percent light re-emitted versus light lost as heat—above 90% in both solution and solid. Caption. Credit: Jinsang Kim Laboratory.“TGlu was an intermediate step for another chemical design, but during purification I found it was surprisingly highly emissive, not only in solution but also in solid state,” Heo said.&nbsp;The discovery led to the systematic study to establish the optimal design. The result was a simple design: a single benzene ring core—six carbon atoms joined in a hexagon. The researchers positioned two groups that give away electrons, called donor groups, across the ring from one another. Next to the donors, they placed two acceptor groups, which withdraw electrons, across the ring from one another.&nbsp;“This so-called quadrupolar structure symmetrically distributes charge across the molecule, providing stable emission in various environments,” Heo said.Because the ring has only six points, donor and acceptor groups are positioned next to each other. This spatial arrangement reduces the energy gap compared to other similar molecules within a compact framework, which means the fluorophore needs a relatively small amount of energy to move an electron from the ground state to an excited state—similar to jumping up a rung on a ladder. &nbsp;However, the molecule’s small size means overall conjugation length remains limited—meaning electrons cannot spread out too far across the molecule. This keeps the absolute energy gap—the distance between ladder rungs—wide enough to emit blue light instead of shifting towards narrower energy gap colors like red.&nbsp;Typically, small band gaps come with an efficiency drawback. When in the excited state on the higher rung of the ladder, an electron can either emit light as it comes back down to the ground state or lose energy as heat through vibration. Often, small band gaps mean more heat loss, reducing the quantum yield—an efficiency metric expressed as the percentage of absorbed UV light that gets reemitted as visible light relative to the amount lost as heat.&nbsp;After trying a series of acceptor groups, the researchers found one that stabilizes the excited state. Even with the small band gap, this acceptor group prevents heat loss by restricting access to what are known as conical intersections, which function as “exit doors” for energy leakage. This unexpected behavior, called an Inverted Energy Gap Law, was confirmed both by experiments and quantum chemical simulations.In the solid state, the acceptor groups, which were intentionally designed to be bulky, prevent the molecules from getting too close to one another which causes fluorophores to lose brightness as energy escapes as heat instead of light, a phenomenon known as quenching.The small, highly-efficient fluorophore is simple to produce—only requiring three steps—which increases its scalability while reducing production costs.The current TGlu design fluoresces blue light. As next steps, the researchers will adjust the band gap, and thus the color. Further, while a high quantum yield from light excitation is promising, device performance under electrical excitation requires separate testing due to additional loss mechanisms. Heo also plans to work toward a phosphorescent version of the molecule, as phosphors are overall more energy-efficient than fluorophores, for use in display technology.Autonomous University of Madrid, University of Valencia, Eberhard Karls University Tübingen and Seoul National University also contributed to this research.&nbsp;

Reaching 98% efficiency in a solid state and 94% in solution, the small fluorescent molecule's design could cut down development time and cost for future applications.

Displays, imaging and sensing: New blue fluorophore breaks efficiency records in both solids and solutions

Among emerging solutions, Direct Retinal Projection (DRP) stands out. Rather than projecting onto a screen, DRP beams images directly onto the retina to unlock ultra-clear visuals with minimal eye strain. Yet the effectiveness of DRP hinges on the performance of the underlying light modulation device. That’s why TDK is now repurposing thin film lithium niobate (TFLN), a material long prized in telecommunications, for next-generation AR smart glasses displays.For decades, lithium niobate has been trusted for high-speed modulation in long-haul infrared optical communication. TDK's breakthrough lies in applying this well-established material to the visible light spectrum and adapting it for DRP imaging. With this world-first innovation, TDK is delivering on promises of higher resolution and significantly lower power consumption.<h2>The TFLN Device Concept</h2>In traditional laser projection systems, color modulation is achieved by dynamically adjusting the injection current to red, green, and blue (RGB) laser diodes. This current-based approach alters the power level of each laser, creating the desired mix of colors. However, this method introduces two fundamental constraints: significant power consumption due to current control to change colors thermal inefficiencies and a practical upper limit in switching speed (i.e., typically around 1 GHz) due to the carrier dynamics in semiconductor laser junctions.TDK’s lithium niobate-based projection module fundamentally rewrites this equation. Instead of modulating color through current, the system fixes the injection current at a low, stable level and applies a voltage across a TFLN electro-optic modulator to dynamically alter the phase of each RGB laser beam. This separation of power delivery (via fixed current) and modulation (via voltage) brings profound improvements in both efficiency and speed.At the center of the system is an Active Photonic Integrated Circuit (Active-PIC) fabricated from sputtered TFLN. These modulators leverage the Pockels effect, which is an intrinsic property of lithium niobate wherein the refractive index changes linearly with an applied electric field. By exploiting this effect, TDK can induce precise phase shifts in the optical waveguides without generating excess heat or requiring large drive currents. The result is color modulation at frequencies up to 10 GHz and more, a number ten times faster than what’s possible with laser diode current modulation.In simple terms, TFLN enables color tuning without altering the energy driving the laser itself. This stability makes it ideal for the DRP method, as laser beams must scan across the retina at high speeds to build up an image. Faster switching means higher effective resolution and smoother rendering, while lower power draw extends battery life in AR smart glasses.<h2>TDK’s Breakthrough in Sputtered TFLN</h2>One of the primary challenges that historically prevented lithium niobate from entering the consumer electronics space was its fabrication method. Traditionally, TFLN devices have been manufactured by bonding thick, bulk-grown crystals onto silicon or other substrates. While this method works well for high-performance optical communication systems, it presents serious drawbacks for mass-market applications. Bulk bonding is inherently low-yield, expensive, and difficult to scale. It also limits device miniaturization due to the thickness and fragility of the bonded crystals.TDK’s breakthrough comes in the form of a proprietary sputtering process that enables the deposition of high-quality, single-crystal-like TFLN directly onto sapphire substrate. Using a precisely tuned sputter deposition technique, TDK forms several hundred nanometer-thick TFLN layers with uniform thickness and crystallinity across large-diameter wafers. This process is compatible with standard semiconductor fabrication workflows and allows wafer-scale integration alongside etching, lithography, and metallization steps.Bonded films require manual alignment and polishing. In contracts, sputtered TFLN can be patterned and processed to drastically reduce cost and manufacturing complexity. Furthermore, TDK’s sputtered films exhibit excellent electro-optic performance, with X-ray diffraction confirming the good crystallinity required for low-loss optical waveguides. This has allowed TDK to fabricate fully functional phase modulators and multiplexers in TFLN for the visible light spectrum: a world-first.<h2>A World-First in Visible Light Modulation</h2>While lithium niobate has long been used in infrared applications, TDK is the first to demonstrate its viability for full-color visible light modulation. This includes independent modulation of red, green, and blue lasers and real-time color mixing through an integrated multiplexer.In lab demonstrations, TDK has successfully shown RGB phase modulation and full-color output using their sputtered-film Active-PIC. Notably, TDK has also proved that the system works within the constraints of wearable AR smart glasses validating the design with direct retinal projection demos. <h2>Three Advantages of TDK’s TFLN</h2><h3>Modulation Speed</h3>In traditional laser-based systems, modulation speed is limited by the physical properties of the laser diode itself. Specifically, the carrier recombination rate within the semiconductor junction caps switching frequencies around 1 GHz. While sufficient for lower-resolution displays, this frequency limit restricts how quickly the laser beam can toggle between red, green, and blue during raster scanning. For DRP applications where individual laser pulses trace an image line by line across the retina, this modulation rate becomes a bottleneck to achieving higher resolution.TDK’s TFLN modulators bypass this limitation by decoupling color control from laser current. Instead, they leverage the Pockels effect to enable optical phase modulation purely through applied voltage at frequencies up to 10 GHz and more. This tenfold increase directly translates into faster pixel transitions and a resultant higher number of color changes per scan line and per frame. Practically, this means TFLN modulators can support 4K and 8K retinal imaging while maintaining frame rates up to 60 fps without flicker.  It also opens the door to future high-frame-rate, high-pixel-density displays such as those required for light-field rendering and retinal microdisplays.<h3>Scalability</h3>The scalability of TDK’s TFLN architecture stems from its use of Active-PIC design. Whereas discrete optical systems require manual assembly of modulators, multiplexers, and beam combiners, the Active-PIC consolidates all light-processing elements onto a single TFLN chip. Each waveguide, multiplexer, and modulator is lithographically defined with submicron precision, enabling complex optical functionality in a compact footprint.Such an approach allows resolution to scale through architectural duplication, not through physical expansion. For example, increasing from 720p to 4K+ is as simple as expanding the number of modulator paths and multiplexing channels on the chip, without enlarging the overall module or increasing system power. The uniformity and reproducibility of the sputtered TFLN process ensure consistent performance across high-density arrays for future displays to be scalble. Because all modulation is done on-chip, TDK’s system also reduces optical alignment complexity and allows for cohesive integration with advanced MEMS mirrors and other beam-steering technologies.<h3>Efficiency</h3>Efficiency in optical systems is largely governed by how light is generated and modulated. Conventional current-driven laser modulation requires constant active power to vary intensity, often resulting in unnecessary thermal load and significant energy loss due to resistive and thermal inefficiencies in the diode junctions. In AR glasses, this translates to shorter battery life, increased heat near the user’s face, and thicker device form factors to accommodate thermal management.TDK’s use of TFLN reverses this paradigm. With the injection current fixed near the diode’s threshold and modulation achieved through external voltage, the optical output remains highly stable while power consumption is minimized. Voltage modulation in TFLN requires much less energy per switch and introduces negligible thermal dissipation, even at GHz speeds. In practical terms, this architecture consumes up to 75% less power than traditional current-modulated systems.This has cascading benefits: battery size can be reduced, wearables stay cooler to the touch, and thermal drift, which can degrade image fidelity, is minimized. Furthermore, low-power operation unlocks more sustainable AR systems that are viable for extended wear.<h2>TFLN Imaging Beyond AR Smart Glasses</h2>Although AR smart glasses is the initial target, the implications of TDK's TFLN extend well beyond. Its high-speed modulation and low-power characteristics make it well-suited for<ul><li>Data Center Optical Interconnects: Lower power and faster IO for photonic computing. Specifically, this is more important for a generative AI data center.</li><li>Quantum Photonics: TFLN's precise phase control and electro-optic properties are suitable for quantum circuits.</li></ul>TDK's work with visible light also unlocks new possibilities for consumer-level Visible Light Communication (VLC) and ultrafast wireless transfer, which have long been limited by cost and complexity.<h2>Scaling and Industry Adoption</h2>As with any material breakthrough, industry adoption depends on cost, scalability, and ecosystem readiness. TDK’s sputtered TFLN approach addresses all three. Combining TFLN modulators with photodetectors, amplifiers, and logic layers could enable fully self-contained one-chip ultracompact optical transceiver to be suitable for optical interconnect for generative AI.<h2>Conclusion</h2>TDK's use of TFLN for visible light modulation is a major moment in the evolution of AR smart glasses display systems. By replacing traditional current-based laser control with high-speed, voltage-driven phase modulation, TDK has enabled higher resolution, lower power consumption, and unprecedented scalability.To learn more, <a href="https://www.tdk.com/en/contact/form/index.html?product_information%255binquiry_classification%255d=ads&amp;product_information%255bproduct_category_1%255d=ads-arvr&amp;product_information%255bproduct_category_2%255d=ads-arvr-001" target="_blank" rel="noopener noreferrer">get in touch with TDK here.</a><a href="https://www.wevolver.com/profile/tdk.corporation"></a>

As augmented smart glasses technology edges closer to mainstream adoption, engineers face challenges around display performance. Namely, many seek to achieve higher resolution without driving up power consumption and device bulk.

image processing

ORGANIZATIONS.

SHAPING THE INDUSTRY.

Latest Posts

Re-Architecting Additive Manufacturing for Scale

How Hyper-detailed Cameras Will Make Recycling More Efficient

Combining Language and Vision – Looking at CLIP

Displays, imaging and sensing: New blue fluorophore breaks efficiency records in both solids and solutions

World-First Adoption of Thin Film Lithium Niobate (TFLN) as Imaging Device