Perspective: Towards sustainable exploration of chemical spaces with machine learning
Authors: Leonardo Medrano Sandonas, David Balcells, Anton Bochkarev, Jacqueline M. Cole, Volker L. Deringer, Werner Dobrautz, Adrian Ehrenhofer, Thorben Frank, Pascal Friederich, Rico Friedrich, Janine George, Luca Ghiringhelli, Alejandra Hinostroza Caldas, Veronika Juraskova, Hannes Kneiding, Yury Lysogorskiy, Johannes T. Margraf, Hanna Türk, Anatole von Lilienfeld, Milica Todorović, Alexandre Tkatchenko, Mariana Rossi, Gianaurelio Cuniberti
Organizations: Institute for Materials Science and Max Bergmann Center of Biomaterials, TUD Dresden University of Technology, 01062, Dresden, Germany · Hylleraas Centre for Quantum Molecular Sciences, Department of Chemistry, University of Oslo, P.O. Box 1033, Blindern, 0315 Oslo, Norway · Interdisciplinary Centre for Advanced Materials Simulation (ICAMS), Ruhr-University Bochum, 44780 Bochum, Germany · Ray Dolby Centre, Cavendish Laboratory, Department of Physics, University of Cambridge, J. J. Thomson Avenue, Cambridge, CB3 0US. United Kingdom · Inorganic Chemistry Laboratory, Department of Chemistry, University of Oxford, Oxford OX1 3QR, United Kingdom · Center for Advanced Systems Understanding, Helmholtz-Zentrum Dresden-Rossendorf, Germany · Center for Scalable Data Analytics and Artificial Intelligence Dresden/Leipzig, TU Dresden, Germany · Institute of Solid Mechanics, TUD Dresden University of Technology, George-Bähr-Str. 3c, 01069 Germany · Machine Learning Group, Technische Universität Berlin, 10587 Berlin, Germany · Institute of Nanotechnology, Karlsruhe Institute of Technology, Kaiserstr. 12, 76131 Karlsruhe, Germany. · Theoretical Chemistry, TUD Dresden University of Technology, Germany · Institute of Ion Beam Physics and Materials Research, Helmholtz-Zentrum Dresden-Rossendorf, Germany · Federal Institute for Materials Research and Testing (BAM), Unter den Eichen 87, 12205 Berlin, Germany · Friedrich-Schiller-University Jena, Institute of Condensed Matter Theory and Optics, Max-Wien-Platz 1, 07743 Jena, Germany · Scientific Computing Center, Karlsruhe Institute of Technology, 76131 Karlsruhe, Germany · Faculty of Computer Science, TUD Dresden University of Technology, 01062 Dresden, Germany · Chemistry Research Laboratory, Department of Chemistry, University of Oxford, Oxford, OX1 3TA, United Kingdom · University of Bayreuth, Bavarian Center for Battery Technology (BayBatt), Bayreuth, Germany · Ecole Polytechnique Fédérale de Lausanne, Institute of Materials, Lausanne, 1015 Switzerland · Department of Chemistry, Chemical Physics Theory Group, University of Toronto, St. George Campus, Toronto, ON M5R 0A3, Canada · Department of Mechanical and Materials Engineering, University of Turku, FI-20014 Turku, Finland · Department of Physics and Materials Science, University of Luxembourg, L-1511 Luxembourg City, Luxembourg · MPI for the Structure and Dynamics of Matter, 22761 Hamburg, Germany · Yusuf Hamied Department of Chemistry, Lensfield Road, Cambridge CB2 1EW, United Kingdom · Cluster of Excellence CARE, TU Dresden and RWTH Aachen, Germany · Cluster of Excellence CeTI, TU Dresden, Germany
Artificial intelligence is transforming molecular and materials science, but its growing computational and data demands raise critical sustainability challenges. In this Perspective, we examine resource considerations across the AI-driven discovery pipeline--from quantum-mechanical (QM) data generation and model training to automated, self-driving research workflows--building on discussions from the ``SusML workshop: Towards sustainable exploration of chemical spaces with machine learning'' held in Dresden, Germany. In this context, the availability of large quantum datasets has enabled rigorous benchmarking and rapid methodological progress, while also incurring substantial energy and infrastructure costs. We highlight emerging strategies to enhance efficiency, including general-purpose machine learning (ML) models, multi-fidelity approaches, model distillation, and active learning. Moreover, incorporating physics-based constraints within hierarchical workflows, where fast ML surrogates are applied broadly and high-accuracy QM methods are used selectively, can further optimize resource use without compromising reliability. Equally important is bridging the gap between idealized computational predictions and real-world conditions by accounting for synthesizability and multi-objective design criteria, which is essential for practical impact. Finally, we argue that sustainable progress will rely on open data and models, reusable workflows, and domain-specific AI systems that maximize scientific value per unit of computation, enabling efficient and responsible discovery of technological materials and therapeutics.
Figures & tables
Figure 1: Scheme illustrating the sustainability topics discussed at the SusML workshop (Dresden, Germany) and in this Perspective, spanning the AI-driven discovery pipeline from quantum-mechanical (QM) data generation and model training to automated, self-driving research workflows.
Figure 2: Standard workflow for developing predictive models of molecular and materials properties, combining quantum-inspired representations with machine-learning (ML) techniques ( e.g. , neural networks, kernel methods, and tree-based models). Certain ML models can be made interpretable through additional explainability techniques, such as symbolic regression, attention mechanisms, and SHAP (SHapley Additive exPlanations) analysis..
Figure 3: Data-driven discovery of traditional vs . emerging 2D materials. ( a ) 2D materials such as graphene are traditionally exfoliated from naturally layered vdW compounds such as graphite. Large computational materials databases ( b ) have already fueled the discovery of several thousand such systems. For the emerging class of non-vdW 2D materials derived from non-layered crystals such as hematite ( α -Fe 2 O 3 ) ( c ), data-driven research and ML based modelling is still at an early stage. Colors: C brown, Fe gray, and O red.
Figure 4: General scheme of structure generation with generative models. Generative models are trained to learn the underlying distribution of a dataset of molecules or materials encoded in a representation, together with their conditioning label on desired properties (gray arrows). From a trained model, new data with the desired properties can be sampled from noise, along with the respective conditioning (blue arrows). The included histogram is reproduced from Ref. [ [ 401 ] ].
Figure 5: Classical and AI-powered approaches along the PSPP relationship [ 111 ] : Classical approaches (i) to connect the different parts range from quantum theory and molecular dynamics approaches up to continuum models. The AI-powered PSPP-chain reasoning process (ii) offers an alternative approach, where every connection is represented by a data-driven model. Models can also leapfrog steps, such as Processing-Property models [ 426 ] .
Figure 6: Conceptual overview of sustainable extractive language models for domain-specific applications in materials science: (a) examples of published domain-specific BERT models derived from structured datasets across different domains (e.g., batteries, photocatalysis, thermoelectrics, photovoltaics, and stress–strain properties); (b) conventional training process for a domain-specific language (BERT) model; (c) resource-efficient creation of domain-specific models via knowledge distillation; (d) integration of small, domain-specific BERT models as agents within an agentic AI workflow for data-driven materials discovery.
Figure 7: Data efficiency for machine-learned interatomic potential (MLIP) models. This schematic overview summarises two aspects, building upon a talk at the SusML workshop. The first aspect ( yellow ) is being efficient in the selection of training data; this comprises ( a ) the exploration of configurational space with evolving MLIP models and ( b ) the use of automated workflows. The second aspect ( blue ) is being efficient in the way that the models are fitted. The pre-training of graph-based models, and their subsequent fine-tuning on much smaller amounts of data (panel c ), is particularly timely in the age of large general-purpose (also referred to as ‘foundational’) MLIPs. Furthermore, evaluation speed at runtime can be improved by model distillation (panel d ): the sketch on the right illustrates a data-driven teacher–student approach [ 287 ] which uses an accurate, but slow model to create training data and fit a more specialised, much faster model to those. The drawing at the top is similar to one in a recent Comment article, first published in Ref. [ [ 40 ] ] by Springer Nature. Panels a and b are adapted from Ref. [ [ 260 ] ], which is published under a CC BY licence (http://creativecommons.org/licenses/by/4.0/).
Figure 8: Scheme illustrating the desired workflow for sustainable exploration of chemical space in the discovery of technological materials and therapeutics. The process begins with the definition of a scientific question, followed by the navigation of existing big data repositories and the reuse of computational workflows and general-purpose models through collaborative partnerships. When implemented correctly, this approach can culminate in the development of self-driving laboratories capable of addressing the original question at larger scales.
While large language models (LLMs) have transformed AI agents into proficient executors of computational materials science, performing a hundred simulations does not make a researcher. What distinguishes research from routine execution is the progressive accumulation of knowledge - learning which approaches fail, recognizing patterns across systems, and applying understanding to new problems. However, the prevailing paradigm in AI-driven computational science treats each execution in isolation, largely discarding hard-won insights between runs. Here we present QMatSuite, an open-source platform closing this gap. Agents record findings with full provenance, retrieve knowledge before new calculations, and in dedicated reflection sessions correct erroneous findings and synthesize observations into cross-compound patterns. In benchmarks on a six-step quantum-mechanical simulation workflow, accumulated knowledge reduces reasoning overhead by 67% and improves accuracy from 47% to 3% deviation from literature - and when transferred to an unfamiliar material, achieves 1% deviation with zero pipeline failures.
Haonan Huang
Department of Physics, Princeton University, Princeton, NJ 08540, USA.
Machine learning has accelerated quantum chemistry but is hindered by the prohibitive cost of generating high fidelity training data. Multifidelity machine learning (MFML) mitigates this overhead by systematically combining abundant low fidelity data with sparse high fidelity data. In spite of its success, standard MFML schemes rely on pre-defined scaling factors to determine sparse data ratio across fidelities, often generating redundant multifidelity data resulting in a loss of efficiency. Here, we introduce an adaptive on-the-fly multifidelity framework for machine learning that autonomously determines training dataset composition. By dynamically querying training samples at each fidelity, the algorithm saturates model accuracy at lower fidelities before moving up to more expensive reference calculations. We benchmark the novel adaptive-MFML across diverse chemical properties including the computational chemistry gold standard coupled cluster energies, and the more chemically challenging excitation energies. In our numerical experiments we show that our adaptive algorithm reduces data generation costs by up to a factor of 30 compared to single fidelity methods and improves upon standard MFML by up to a factor of 5. The mitigation of data redundancy establishes a high-accuracy low-cost pathway for sustainable cost-aware machine learning in quantum chemistry.
Vivin Vinod, Peter Zaspel
School of Mathematics and Natural Sciences, University of Wuppertal, Germany.
The rapid maturation of artificial intelligence (AI) and machine learning (ML) has catalyzed a profound shift in how chemical engineering problems are formulated, analyzed, and solved. Advances in computing, data availability, and learning algorithms have enabled AI/ML methods to impact applications spanning atomic-scale simulations, materials and catalyst discovery, transport and thermodynamics, separations, process systems engineering, and industrial operations. This article provides a perspective on recent methodological developments and representative applications, emphasizing how AI/ML tools are being integrated with first-principles models to address challenges of predictive accuracy, data scarcity, extrapolation, interpretability, and model lifecycle management. Across domains, a unifying trend is the move away from purely black-box approaches toward hybrid and physics-informed frameworks that explicitly respect conservation laws, thermodynamic consistency, and known structural constraints. These approaches not only improve robustness and reliability, but also enable meaningful human-AI collaboration by providing information at an appropriate level of abstraction for the task and decision context. We conclude that AI and ML are not replacing the core principles of chemical engineering; rather, they are amplifying them. As the field advances toward increasingly autonomous, adaptive, and sustainable systems, the thoughtful integration of AI/ML with first-principles understanding and domain expertise will be essential to realizing their full potential across both research and industrial practice.
Michael Baldea, Linda J. Broadbelt, Marianthi G. Ierapetritou +17
McKetta Department of Chemical Engineering and Oden Institute for Chemical Engineering and Sciences, The University of Texas at Austin, Austin, TX 78731, USA · Department of Chemical and Biological Engineering, Northwestern University, Evanston, IL 60208, USA · Department of Chemical and Biomolecular Engineering, University of Delaware, Newark, DE 19716, USA +13