Open Access

Despite significant progress, computational materials design still faces major challenges—particularly when simulating advanced and chemically complex materials with the accuracy of density functional theory (DFT) or beyond.[1] To overcome these limitations, machine learning (ML) methods have gained considerable traction in recent years. We have developed robust data-generation strategies to support the creation and benchmarking of new ML models.[2] In this talk, I will highlight methods for large-scale quantum-chemical bonding analysis and workflows for ML interatomic potentials. Our work demonstrates that quantum-chemical bonding properties can be incorporated into ML models to predict phononic properties.[3] This approach enables large-scale validation of expected correlations—such as the link between bonding strength and force constants or thermal conductivities. Furthermore, we have built an automated training framework for machine-learned interatomic potentials (autoplex).[4] Initial workflows include random structure searches, suitable for general-purpose potentials, as well as specialized workflows targeting ML potentials with accurate phonon properties. While atomistic simulations are highly effective for certain material properties, others—such as magnetism or synthesizability—remain challenging. In these cases, promising strategies include benchmarking established ab initio methods against chemical heuristics or developing new ML models primarily based on experimental data.[5,6]

Machine learning (ML) offers new routes to overcome the limitations of density functional theory (DFT) for advanced materials. We present data-generation strategies and workflows for ML interatomic potentials, including large-scale quantum-chemical bonding analysis.[1,2,3] Incorporating bonding descriptors into ML models enables prediction of phononic properties and validation of correlations between bonding strength, force constants, and thermal conductivity.[3] We introduce autoplex, an automated framework for training ML potentials, supporting general-purpose and phonon-focused workflows.[4] These developments provide a basis for fine-tuning foundation models for thermal transport at reduced cost.[5] For properties such as magnetism or synthesizability, we discuss complementary approaches, comparing ab initio methods with chemical heuristics and experimental data-driven ML models.[6,7]Our work advances scalable, accurate simulations for materials discovery.

Machine learning (ML) offers powerful new strategies for accelerating the discovery and design of functional materials. In our work, we develop ML models and software frameworks for large-scale screening and advanced materials simulations, starting from robust high throughput quantum chemical workflows, such as those implemented in atomate2.[1,2] These automated workflows enable the creation of large, high quality materials databases that form the foundation for data science and machine learning. In addition to experimentally known crystal structures, increasingly generative models are used to extend materials databases, which also need to be evaluated.[3] To build predictive, scientifically grounded ML models, we use chemical bonding concepts, incorporating quantum chemical bonding strengths and related descriptors as physically meaningful features to predict vibrational properties and heat transport.[4,5] Beyond property prediction, we address the challenge of determining which hypothetical materials are synthesizable. To this end, we introduced co-training into a positive–unlabelled (PU) learning framework, enabling ML based classification even in the absence of true negative data—an essential step for screening synthesizable compounds.[6,7] To advance atomistic simulations of complex materials, we further developed automated training pipelines for ML interatomic potentials that support both general-purpose and system specific potential development, as implemented in our software autoplex.[8] This automated approach has already facilitated detailed investigations of challenging systems, including the computational exploration of amorphous arsenic.[9] Together, these developments provide a toolbox spanning workflow automation, automated ML potential training, and ML models for materials properties and synthesis, enabling scalable, data driven discovery and understanding of advanced materials.

Machine learning (ML) offers new routes to overcome the limitations of density functional theory (DFT) for advanced materials. We present data-generation strategies and workflows for ML models, including large-scale quantum-chemical bonding analysis and ML interatomic potentials. Incorporating bonding descriptors into ML models enables the prediction of phononic properties and the validation of correlations between bonding strength, force constants, and thermal conductivity. We further introduce the software package autoplex, an automated framework for training ML potentials, supporting general-purpose and phonon-focused workflows. These developments additionally provide a basis for fine-tuning foundation models for thermal transport at reduced cost. For properties such as magnetism or synthesizability, we discuss complementary approaches comparing ab initio methods with chemical heuristics and experimental data-driven ML models. Our work advances scalable, accurate simulations for materials discovery.

My talk covered, among other things, robust data generation for machine learning. It showed how heuristics can be used within machine learning models and how they might also be extracted from machine learning models. Beyond this, I showed an automated pipeline for training machine learning potentials.