Kevin Maik Jablonka

Predicting properties from coordinate–category data—sets of vectors paired with categorical information—is fundamental to computational science, manifesting in materials science as predicting formation energies or elastic moduli from crystal structures. While large language models (LLMs) have increasingly been applied to such tasks, optimal strategies for achieving reliable predictions remain elusive. Here we report fundamental limitations in LLM’s ability to learn from coordinate information in coordinate–category data. Through systematic experiments using synthetic datasets with tunable coordinate and category contributions, and a comprehensive benchmarking framework (MatText), we find that LLMs consistently fail to capture coordinate information while excelling at category patterns. This geometric blindness persists regardless of model size (up to 70 billion parameters), dataset scale (up to 2 × 10 ⁶ structures) or text representation strategy. For material property prediction tasks dominated by structural effects, specialized geometric architectures consistently outperform LLMs, as evidenced by a clear ‘Graph Neural Network Language Model (GNN-LM) wall’ in performance benchmarks. Our analysis provides concrete guidelines for architecture selection in scientific machine learning.

Materials science is at the crossroad between fundamental and applied sciences. Whether enabling clean energy, next-generation computing, or advanced manufacturing, it shapes the tools we use and the systems we build. As our societies undergo rapid digital, environmental, and technological transitions, materials science becomes even more central. It's a space where innovation can respond to practical challenges while aligning with broader social values. In the European context, that means supporting sustainability, openness, and solidarity -while also strengthening competitiveness.This roadmap explores how digital tools-especially simulation, data science, and AI-are transforming materials research, in connection with the twin digital and energy transition. The digital transition refers to the widespread adoption of digital technologies and data-driven methods across sectors, while the green (or energy) transition focuses on shifting toward sustainable, lowcarbon energy systems-together forming what is often called the twin transition, a joint effort to make economies both smarter and more sustainable.The Roadmap outlines both the technical directions and the cultural shifts needed to make this transformation inclusive and effective. Chapters span from atomic-scale simulations to advanced experimentation, from reproducibility to intelligent optimization, and from institutional reform to education for the next generation.Importantly, this isn't a single viewpoint. The document brings together a wide range of voices: researchers from different disciplines, working across length and time scales. It includes early-career scientists and senior experts. Enabled by activities supported by European Cooperation in Science and Technology (COST), it reflects a commitment to gender and geographic diversity. This plurality doesn't just enrich the content-it makes the vision more robust and relevant.We hope this collection serves not just as a guide, but as an invitation to collaborate-across fields, sectors, and borders-as we reimagine the future of materials science in a digital age.

Naming molecules systematically remains an open challenge in cheminformatics. The reverse task—converting names to structure—has been largely solved for over a decade by OPSIN, yet no open-source tool performs the forward direction. Here we introduce Openclatura, which generates IUPAC names from molecular structures. The rule set was constructed by an agentic system that exhaustively enumerates naming rules, with each candidate name checked for structural round-trip consistency against OPSIN. Across three challenging benchmarks (QM9, PubChem, and ZINC22), Openclatura produces valid names for 98% of molecules, compared to 95% for the deep-learning baseline STOUT, while running on a single CPU core (40 times faster). We further present Robomoleculographer, an extension that generates natural-language descriptions of molecular graphs, bringing to chemistry what robocrystallographer offers to materials. Openclatura is fully deterministic and exposes an inspectable decision trace for every generated name.

Chemists often use different formats and modalities to describe the same polymers (e.g., PSMILES, BigSMILES, systematic names), molecules (e.g., SMILES, 1H NMR, IR), and materials (e.g., MOFid, pXRD, crystal structures). This fragmentation makes it challenging to leverage all available data using classical machine learning approaches. To address this, we developed ChemFuse, a contrastive learning-based approach to align fragmented data representations into a shared embedding space. ChemFuse’s aligned embeddings allow us to combine data from multiple sources to improve downstream predictive models by increasing the amount and diversity of the training data. The model performance improves by 44.7% for polymers (aligning PSMILES and BigSMILES), 44.7% for molecules (aligning SMILES and 1H NMR), and 44.6% for MOFs (aligning

Naming molecules systematically remains an open challenge in cheminformatics. The reverse task—converting names to structure—has been largely solved for over a decade by OPSIN, yet no open-source tool performs the forward direction. Here we introduce Openclatura, which generates IUPAC names from molecular structures. The rule set was constructed by an agentic system that exhaustively enumerates naming rules, with each candidate name checked for structural round-trip consistency against OPSIN. Across three challenging benchmarks (QM9, PubChem, and ZINC22), Openclatura produces valid names for 98% of molecules, compared to 95% for the deep-learning baseline STOUT, while running on a single CPU core (40 times faster). We further present Robomoleculographer, an extension that generates natural-language descriptions of molecular graphs, bringing to chemistry what robocrystallographer offers to materials. Openclatura is fully deterministic and exposes an inspectable decision trace for every generated name.

Elucidating molecular structures from spectroscopic data remains one of chemistry’s most fundamental challenges, typically requiring extensive expert knowledge and manual interpretation of multiple analytical techniques. This is because the structure elucidation problem often has degenerate solutions for a limited set of experimental data. Existing computational approaches are limited to single spectroscopic modalities, require extensive manual preprocessing, and lack the confidence estimates and context necessary for practical application. Here we present , a framework that combines contrastive learning with evolutionary algorithms to automate structure elucidation directly from raw, multimodal spectroscopic data. By aligning embeddings across NMR, infrared, and mass spectrometry, mimics how experts use multiple spectroscopic lenses while providing calibrated confidence scores and relevant database context. On challenging molecular identification tasks, matches expert chemist performance in head-to-head comparisons in a pilot study. The system successfully identifies incorrect structure assignments in published literature and adapts to new chemical domains without retraining by updating its reference database. Our approach demonstrates how synergistic combination of machine learning paradigms can solve analytical bottlenecks that have constrained chemical discovery.

The architecture of a copolymer determines how the material behaves. Designing a polymer for a target property, therefore, should begin with designing for a target architecture. Yet no method predicts the architecture a given monomer pair will react to under given conditions. Even though reactivity ratios have been known since Mayo and Lewis to depend on solvent, temperature, and polymerization mechanism, every predictive approach, from the Q–e scheme of Alfrey and Price to recent machine-learning models, takes monomer structure alone as input. Measurements for thousands of monomer pairs exist but are scattered across eight decades of literature, much of it predating machine-readable publishing and surviving only as scanned page images. We address this gap with PolyCARP, a classifier that predicts copolymer architecture from monomer pair and reaction conditions before any synthesis, trained on a database we extracted from the literature. A vision–language-model pipeline parses 1,206 publications, yielding 3,791 copolymerizations annotated with reactivity ratios, solvent, temperature, and polymerization mechanism. This, to our knowledge, is the first dataset at this scale to record reaction conditions per entry. For any monomer pair and set of conditions, PolyCARP returns an architecture class and the nearest experimental analogues in the database. We validate the tool against a literature case study, where it captures solvent-driven architectural transitions, and against three prospective laboratory copolymerizations. The database, model, and code are openly released—enabling the community to extend both data and models, and to make more of polymer chemistry predictable.

The architecture of a copolymer determines how the material behaves. Designing a polymer for a target property, therefore, should begin with designing for a target architecture. Yet no method predicts the architecture a given monomer pair will react to under given conditions. Even though reactivity ratios have been known since Mayo and Lewis to depend on solvent, temperature, and polymerization mechanism, every predictive approach, from the Q–e scheme of Alfrey and Price to recent machine-learning models, takes monomer structure alone as input. Measurements for thousands of monomer pairs exist but are scattered across eight decades of literature, much of it predating machine-readable publishing and surviving only as scanned page images. We address this gap with PolyCARP, a classifier that predicts copolymer architecture from monomer pair and reaction conditions before any synthesis, trained on a database we extracted from the literature. A vision–language-model pipeline parses 1,206 publications, yielding 3,791 copolymerizations annotated with reactivity ratios, solvent, temperature, and polymerization mechanism. This, to our knowledge, is the first dataset at this scale to record reaction conditions per entry. For any monomer pair and set of conditions, PolyCARP returns an architecture class and the nearest experimental analogues in the database. We validate the tool against a literature case study, where it captures solvent-driven architectural transitions, and against three prospective laboratory copolymerizations. The database, model, and code are openly released—enabling the community to extend both data and models, and to make more of polymer chemistry predictable.

Machine learning can accelerate materials discovery. Models perform impressively on many benchmarks. However, strong benchmark performance does not imply that a model learned chemistry. I test a concrete alternative hypothesis: that property prediction can be driven by bibliographic confounding. Across five tasks spanning MOFs (thermal and solvent stability), perovskite solar cells (efficiency), batteries (capacity), and TADF emitters (emission wavelength), models trained on standard chemical descriptors predict author, journal, and publication year well above chance. When these predicted metadata ("bibliographic fingerprints") are used as the sole input to a second model, performance is sometimes competitive with conventional descriptor-based predictors. These results show that many datasets do not rule out non-chemical explanations of success. Progress requires routine falsification tests (e.g., group/time splits and metadata ablations), datasets designed to resist spurious correlations, and explicit separation of two goals: predictive utility versus evidence of chemical understanding.

Data-driven techniques have a large potential to transform and accelerate the chemical sciences. However, chemical sciences also pose the unique challenge of very diverse, small, fuzzy data sets that are difficult to leverage in conventional machine learning approaches. A new class of models, which can be summarized under the term general-purpose models (GPMs) such as large language models, has shown the ability to solve tasks they have not been directly trained on, and to flexibly operate with low amounts of data in different formats. In this review, we discuss the fundamental building principles of GPMs and review recent and emerging applications of those models in the chemical sciences. While many of these applications are still in the prototype phase, we expect that the increasing interest in GPMs will make many of them mature in the coming years.