Materials Discovery Thrust: Five Years of Accelerating Materials Innovation with A

Since the start of the IBM–Illinois Discovery Accelerator Institute, the Materials Discovery Thrust has focused on a simple but ambitious goal: to make the discovery of new materials faster, more systematic, and more useful for real scientific and technological problems.

The work has brought together researchers from the University of Illinois Urbana-Champaign and IBM Research across materials science, chemistry, chemical engineering, computer science, machine learning, molecular simulation, and automation. The common thread has been the development of AI-driven tools that can help scientists move more efficiently from literature and data, to candidate materials, to experimental validation.

Over these five years, the thrust has grown from early efforts in AI-assisted synthesis planning, polymer discovery, and closed-loop experimentation into a broader program around foundation models, scientific workflows, and sustainable materials discovery.

Major Accomplishments

Accelerating Materials Discovery with AI

A central contribution of the thrust has been the development of AI methods that support the full materials discovery process. Early projects addressed the extraction of materials information from the literature, virtual screening of candidates, retrosynthesis planning, and the use of automated laboratories for synthesis and characterization.

One important direction was the development of AI tools for synthesis planning. The thrust supported work on chemoenzymatic retrosynthesis, including ACERetro, a method designed to combine chemical and enzymatic synthesis logic and identify more efficient synthetic routes. This work is relevant because making a material is often as challenging as identifying it computationally. Better synthesis planning helps close the gap between design and practice.

Another important direction was closed-loop discovery for organic photovoltaic materials. In this work, automated synthesis, machine learning, Bayesian optimization, and experimental testing were combined to identify molecular motifs with improved photostability. This demonstrated the value of bringing AI directly into the experimental cycle, where each round of results can guide the next one.

The thrust also moved into foundation models for materials science. Projects explored foundation AI models for inorganic crystalline materials, polymeric materials, and chemical synthesis. One example is FaithRetro, an LLM-based agent for retrosynthesis that can predict precursors, infer reaction conditions, retrieve relevant procedures, and draft experimental protocols. This type of work shows how language models can become practical scientific assistants, provided that they are adapted to the chemistry domain and constrained by reliable data and procedures.

 

Building Data Assets, Platforms, and Workflows

A major part of the thrust’s impact has come from building reusable data resources and computational workflows.

The Open Macromolecular Genome project addressed one of the long-standing bottlenecks in polymer science: the lack of large, accessible, and synthetically meaningful datasets. The project developed a database of 12 million synthesizable polymer chemistries generated from 17 canonical polymerization mechanisms. It also supported the calculation of 25–27 physical properties relevant to polymer design. This work provides a foundation for physics-aware, multi-task models that can help predict and design polymeric materials in a more systematic way.

More recently, the thrust developed a Framework for Accelerated Materials Development, or FAMD, with a focus on sustainable materials. The framework connects literature mining, knowledge extraction, generative molecular design, simulation, predictive modeling, active learning, and experimental validation. This is important because materials discovery is not only a modeling problem. It requires a workflow where information, computation, and experiments continuously inform each other.

The FAMD work was applied to two urgent sustainability problems: rare-earth element separations and PFAS remediation.

For rare-earth separations, the team curated difficult-to-access experimental data, including 2,425 distribution-coefficient measurements across 104 ligands and 7,351 LogK1 entries. These data were used to train predictive models and guide the generation and screening of new ligand candidates. The workflow produced 5,545 valid generated ligand candidates and prioritized experimentally accessible molecules for testing.

For PFAS remediation, molecular dynamics simulations were used to generate interaction labels for fluorophilic copolymer systems. These computational results were then compared with electro-adsorption experiments, creating a first bridge between simulation-based screening and experimental performance for redox-polymer platforms.

 

Connecting AI with Experimental Validation

One aspect that distinguishes the Materials Discovery Thrust is the emphasis on closing the loop between prediction and experiment.

The thrust established experimental workflows for liquid–liquid extraction, ICP-OES quantification of rare-earth elements, and electrochemical testing of redox polymers. In the rare-earth separations work, the team developed a 96-well high-throughput extraction workflow and in-house quantification for six rare-earth elements. In the PFAS work, simulation and electro-adsorption experiments were connected to evaluate polymer systems for environmental remediation.

This connection between AI and experimental validation is essential. It ensures that the models are not only producing plausible candidates on paper, but are also helping to guide real materials decisions in the laboratory.

 

Research Impact and Community Building

The Materials Discovery Thrust has also played an important role in building a collaborative research community between Illinois and IBM.

The projects involved faculty, students, postdoctoral researchers, and IBM scientists working across different disciplines. Regular meetings and technical reviews helped align modeling, simulation, synthesis, and characterization. Graduate students were deeply involved in the work, from data curation and machine learning to molecular simulation and experimental validation.

This training component is one of the most valuable outcomes of the thrust. The next generation of materials researchers will need to be comfortable with both physical science and AI-driven methods. The thrust has helped create exactly this kind of environment.

The thrust also seeded new research directions for the Institute. Early work in automation, polymer discovery, OPVs, and retrosynthesis helped lead to larger efforts in foundation models, polymer databases, inorganic materials modeling, faithful retrosynthesis, and sustainable materials discovery.

By the numbers

  • 12 million synthesizable polymer chemistries represented in the Open Macromolecular Genome.
  • 17 canonical polymerization mechanisms used to construct the OMG chemical space.
  • 25–27 physical properties calculated or targeted to support polymer foundation models.
  • 2,425 curated rare-earth distribution-coefficient measurements across 104 ligands.
  • 7,351 curated LogK1 entries supporting rare-earth ligand modeling.
  • 5,545 valid generated ligand candidates from the rare-earth separations workflow.
  • Approximately 15 experimentally accessible ligands prioritized for testing.
  • Six rare-earth elements quantified through in-house ICP-OES workflows.
  • A 96-well high-throughput liquid–liquid extraction workflow established for experimental screening.
  • Multiple major project lines advanced, including chemoenzymatic retrosynthesis, closed-loop OPV discovery, Open Macromolecular Genome, inorganic materials foundation models, FaithRetro, and AI-driven sustainable materials discovery.

 

Closing Summary

Over its first five years, the Materials Discovery Thrust has helped establish a practical vision for AI-driven materials research. It has produced data resources, models, and workflows for polymer design, synthesis planning, sustainable separations, PFAS remediation, organic photovoltaics, and inorganic materials modeling.

The main achievement is not a single tool or result, but the creation of an integrated way of working: extracting knowledge from scientific literature, using AI and simulation to propose candidates, testing the most promising ones, and feeding the results back into the discovery cycle.

This is the direction in which materials science is moving. The Materials Discovery Thrust has made important progress toward that future, while also building a strong IBM–Illinois research community and training students who can work naturally across AI, computation, chemistry, and experiment.

A schematic showing the Materials Discovery Thrust workflow: literature and patent mining, knowledge extraction, foundation models, molecular generation, simulation, predictive screening, active learning, synthesis, high-throughput characterization, and experimental feedback. This figure would communicate the thrust’s main contribution: transforming materials discovery from a linear trial-and-error process into a closed-loop AI-guided workflow.
A schematic showing the Materials Discovery Thrust workflow: literature and patent mining, knowledge extraction, foundation models, molecular generation, simulation, predictive screening, active learning, synthesis, high-throughput characterization, and experimental feedback. This figure would communicate the thrust’s main contribution: transforming materials discovery from a linear trial-and-error process into a closed-loop AI-guided workflow.