Interview with JMI's Guest Editor: Dr. Huan Tran

Dr. Huan Tran is a Senior Research Scientist at the Georgia Institute of Technology and Director of Research Innovation at Matmerize, Inc., where he develops AI-driven approaches for the discovery and design of advanced materials. He has authored more than 100 scientific publications and contributed to widely used materials informatics methodologies. He serves the materials science community as a journal and grant reviewer, journal editor, and organizer of scientific conferences and symposia. Dr. Tran collaborates closely with academic researchers, national laboratories, and industrial partners to connect computational predictions with experimental validation and accelerate materials innovation and commercialization. His current research interests include generative AI for materials discovery, polymer informatics, and the data-driven design of sustainable, functional polymers.

His profile in LinkedIn is https://www.linkedin.com/in/huan-tran-b1a033443/
Matmerize profile is https://www.linkedin.com/company/matmerize/
Georgia Tech profile is https://www.linkedin.com/school/georgia-institute-of-technology/

Interview Questions and Answers:

Q1. Polymer materials have traditionally relied heavily on experimental trial-and-error approaches. From your perspective, what are the major challenges that limit the efficient discovery and design of advanced polymers, and how can artificial intelligence help overcome these challenges?

Experimental methods have long been predominant and indispensable for the discovery and design of advanced polymeric materials, and I think this situation will not be changed substantially for some time. The fact is that challenges in this area are numerous and enormous, encompassing both technical and non-technical factors that strongly limit the effectiveness of non-experimental alternatives.

The performance/properties of a polymeric material depend not only on its chemical composition and molecular structure but also on its molecular weight, morphology, reinforcements, additives, processing history, operating environment, and other factors that may not be fully understood. The multiple length scales involved in the structural details make comprehensive modeling essentially impractical, while the timescales associated with the relevant chemical and physical processes, the processing procedure, and the operation often extend beyond the reach of current computational capabilities. Despite tremendous efforts that have been made in the history of this area, understanding and evaluating a polymer using physics-based computational methods remain extremely challenging.

Beyond the technical challenges, a polymer with superior properties may not necessarily be adopted by users and market. New materials may be too expensive to produce, difficult to manufacture at scale, or incompatible with existing production facilities. Established materials benefit from mature supply chains, optimized manufacturing processes, and proven reliability. Replacing them can bring in substantial costs, uncertainties, and commercial risks, even with the new materials with superior performance.

AI offers opportunities to address these challenges through rapid property prediction, high-throughput virtual screening, generative design, and multiobjective optimization. By incorporating manufacturing constraints, costs, and sustainability considerations, AI can help identify polymers that are not only technically promising but also commercially viable. Challenges facing AI are also enormous, including, but not limited to, the quality, quantity, diversity, and representativeness of available data. Ultimately, AI cannot replace experiments or overcome all industrial barriers.

I think the most promising direction is an integrated approach combining AI, computational simulations, and targeted experiments, with industrial requirements considered from the earliest design stages. Rather than eliminating experimental trial and error, AI can improve its efficiency and help bridge the gap between scientific discovery and industrial implementation.

Q2. Your research explores the application of machine learning and data-driven approaches in polymer materials design. What are the key factors that determine the success of AI models in predicting polymer properties and guiding material discovery?

AI models rely on data, and thus, are generally bounded by the quality, quantity, diversity, and completeness of their training data. In fact, today’s polymer AI models are constrained more by the lack of well-curated, trustworthy data than by the limitations of algorithms. In my opinion, securing and developing a sufficiently large, diverse, high-quality, and continuously evolving (living) data infrastructure is the most important factor determining the success of AI models in predicting polymer discovery and design. Equally important is how polymer structures, compositions, processing conditions, and other relevant information are represented and transferred into the models. Appropriate representations must capture the essential characteristics governing the properties of interest, while models must recognize deep and hidden patterns in the (represented) data, be validated appropriately to ensure their reliability, particularly when applied to materials outside the training domain.

However, having a powerful and accurate predictive model is not the ultimate objective. It needs to be integrated into a real-world, well-designed, and smoothly functioning materials design loop, in which predictions guide the generation and screening of new candidates, promising materials are selected for experimental validation, and new experimental results are fed back into the database to continuously improve the models. Uncertainty quantification and active learning can make this process more efficient by identifying not only promising candidates but also the most informative experiments. Furthermore, practical constraints such as synthesis feasibility, processing requirements, cost, and scalability should be considered from the beginning rather than after a material has been discovered. Challenges of various kinds can arise at any of these steps, and addressing them is far from straightforward.

Overall, the success of AI in polymer materials design should be measured not simply by predictive accuracy or computational efficiency, but by its ability to integrate into the broader polymer R&D ecosystem, ultimately accelerating the discovery, experimental validation, and practical development of new materials with desirable properties.

Q3. Compared with traditional inorganic materials, polymers often exhibit complex relationships between molecular structures, processing conditions, and final properties. How can AI approaches effectively capture these in polymer systems?

Polymers are highly disordered; thus, their multiscale structure-property relationships are intrinsically complicated. However, polymer backbones, which have no direct counterparts in traditional inorganic materials, offer AI-based approaches some unique opportunities. In fact, many properties of polymers are strongly correlated with the chemical nature and topology of their backbones, where intra-chain interactions are typically much stronger than inter-chain interactions. Rigid backbones containing aromatic rings or double bonds generally restrict chain flexibility, often contributing to high stiffness and strength but limited elongation at break. Conversely, backbones consisting of highly flexible single bonds allow greater chain mobility and deformation, potentially yielding excellent elasticity and toughness. Similar arguments can be made for thermal, rheological, and even certain electrical properties of polymers. Representing polymer backbones using (chain) descriptions such as SMILES or SMARTS is relatively straightforward, and this is one of the key reasons for the effectiveness of AI models based on these simple representations.

Nevertheless, polymer backbone alone is not sufficient to fully describe a polymeric material and its properties. Molecular weight and its distribution, branching, crosslinking, crystallinity, morphology, additives, and processing history can all substantially determine the material properties. To capture these effects, AI models should incorporate descriptors beyond the chemical structure, including but not limited to, molecular characteristics, formulation compositions, processing parameters, and, where available, microstructural information. Different representations and ML architectures can then be employed to capture the nonlinear relationships and interactions among these factors.

An important practical challenge is that such information is often incomplete or unavailable. Models based solely on chemical structures can be effective for predicting certain intrinsic properties but may be less reliable for properties strongly influenced by processing history and morphology. Integrating experimental data, physics-based simulations, and domain knowledge may be helpful for addressing these limitations, and this is the main motivation of physics-informed/enforced models in polymer research.

After all, an optimal strategy is not necessarily always about a fancy and sophisticated AI model, but more about a balanced approach in which potential data deficiencies, in either quality or quantity, are compensated for by domain knowledge and the informed selection and appropriate representation of the most relevant factors governing each target property. A relatively simple model trained on high-quality data with physically meaningful descriptors can often outperform a much more sophisticated model lacking essential information. In summary, capturing polymer structure-property relationships requires a combination of appropriate representations, reliable data, physical understanding, and experimental validation.

Q4. Generative AI and large-scale models are rapidly emerging in materials science. What opportunities do you see for generative models in accelerating polymer discovery, such as designing novel polymer structures with targeted properties?

Generative AI offers exciting opportunities to push polymer discovery and design beyond conventional screening toward inverse design, in which desired properties are specified first and candidate polymer structures are generated. Unlike traditional predictive models that estimate properties from given polymers, generative models propose new molecular structures, monomers, copolymer compositions, and even formulations that likely meet the targets. This capability is particularly valuable given the enormous polymer design space, which is practically impossible to explore exhaustively through conventional computational methods. When coupled with predictive models, generative AI can support iterative design and multiobjective optimization, balancing competing requirements such as mechanical strength, thermal stability, processability, recyclability, and cost. Incorporating chemical constraints and polymerization reaction rules can further improve the feasibility of the generated candidates.

However, generating chemically valid polymer structures is relatively easy compared with ensuring the generated materials can be synthesized, processed, and manufactured at industrial scale. Generative models also inherit the limitations of their training data, and predictions for novel polymers can be highly uncertain. Therefore, uncertainty quantification, synthesis feasibility, and experimental validation are essential components of a reliable generative design framework.

I believe that significant opportunities arise from integrating generative AI with property prediction, physics-based simulations, and targeted experiments into a closed-loop discovery platform. Such an approach can continuously generate, evaluate, validate, and refine polymer candidates, substantially reducing the time and resources required to identify promising materials. Ultimately, the success of generative AI should be measured not by how many novel polymer structures it can generate, but by how effectively it can deliver new materials with desirable properties that can be practically synthesized and processed.

Q5. Data availability and standardization remain important challenges for AI-driven materials research. What strategies do you think are needed to establish reliable polymer databases and AI-ready materials data infrastructures?

Clearly, establishing reliable polymer databases is one of the most important and challenging tasks for AI-driven polymer discovery. A tremendous amount of polymer information has been generated and accumulated over decades in scientific publications, patents, technical datasheets, and industrial laboratories. However, much of this information is fragmented, incomplete, unstructured, inconsistently reported, or inaccessible. Transforming these resources into well-curated, trustworthy, and AI-ready databases is particularly challenging, requiring coordinated efforts among numerous parties with different priorities and agenda.

A reliable strategy is to take advantage of recent advances in natural language processing (NLP) and large language models (LLMs) to extract polymer information from the enormous body of existing literature and technical documents. In our work, we have explored such approaches to build polymer and polymer-composite databases from scientific publications and industrial datasheets. However, automated extraction alone is insufficient. Data must be carefully cleaned, normalized, deduplicated, and validated, ideally with human expertise involved in resolving ambiguous or inconsistent records. Equally important is preserving data provenance so that every entry can be traced back to its original source and experimental context.

Also, data infrastructures should not only be passive repositories. Instead, they should be continuously updated, searchable, and accessible to predictive models, generative design tools, and experimental workflows. In industrial applications, secure mechanisms for incorporating proprietary data are important but not necessarily easy, as some of the most valuable polymer data are held by companies and cannot be publicly shared.

Finally, I believe that greater collaboration among academia, industry, and government agencies is essential. Common data standards, shared benchmarks, and incentives for publishing well-documented experimental results, including unsuccessful experiments, would substantially improve the availability and reliability of polymer data. The final objective should not simply be more data, but to establish a continuously growing, trustworthy data infrastructure that can directly support AI efforts in materials design and experimental validation.

Q6. Looking ahead, how do you envision the future development of AI-driven polymer materials research? What breakthroughs or emerging technologies do you believe will have the greatest impact on accelerating the transition from computational predictions to real-world polymer applications?

I believe the future of AI-driven polymer materials research lies in moving beyond isolated predictive models toward integrated, intelligent materials design platforms that connect data, AI, computational simulations, experiments, and industrial manufacturing. Significant progress has been made in predicting polymer properties and generating novel chemical structures, but the gap between computational predictions and real-world applications remains substantial. Bridging this gap will require not only more powerful AI models but also fundamental changes in how polymer research and development is conducted.

One particularly promising direction is the integration of large language models (LLMs), generative AI, and conventional machine-learning methods into unified materials design frameworks. LLMs can help extract and organize knowledge from scientific literature, patents, and industrial data, while generative models can propose novel polymer structures and formulations with targeted properties. Predictive models can then rapidly evaluate these candidates, considering multiple and often competing requirements, such as mechanical performance, thermal stability, processability, cost, and environmental sustainability. Advances in physics-informed AI and uncertainty quantification will further improve the reliability of these predictions, particularly when exploring materials beyond existing training datasets.

Another important breakthrough will come from autonomous, closed-loop materials discovery, in which AI models guide experiments, experimental results continuously improve the models, and the entire process is optimized through active learning. However, I believe that fully autonomous laboratories are not necessarily the ultimate objective. More important is establishing practical and efficient workflows in which AI and human expertise complement each other, allowing researchers to make better decisions and substantially reduce unnecessary experiments.

From my experience working with industrial partners, I also believe that industrial feasibility must become an integral part of AI-driven materials design rather than an afterthought. A polymer with outstanding predicted properties may still be impractical because of synthesis difficulties, high production costs, limited raw-material availability, or incompatibility with existing manufacturing infrastructure. Future AI platforms should therefore incorporate polymerization chemistry, formulation design, processing conditions, manufacturing constraints, and economic considerations from the earliest stages of discovery. This is particularly important for developing recyclable, biodegradable, and PFAS-free polymers, where performance, sustainability, and commercial viability must be addressed simultaneously.

I envision AI transforming polymer research from a predominantly experimental trial-and-error process into a more systematic, predictive, and increasingly automated approach. The greatest breakthrough will not necessarily be a new AI algorithm or a larger model, but the successful integration of existing and emerging technologies into reliable, accessible, and practically useful materials design systems. Their impact should ultimately be measured by how effectively they accelerate the development and industrial adoption of new polymeric materials, rather than by predictive accuracy or the number of generated candidates alone.

Editor: Eric Luo
Production Editor: Xingyue Luo
Respectfully Submitted by the Editorial Office of Journal of Materials Informatics