Machine-learning techniques for atomistic modeling
Data-driven models enabling accurate and scalable atomistic simulations
Researchers at AMC develop machine-learning methods to accelerate and enhance atomistic simulations of chemical systems. By training models on quantum-mechanical data, we create efficient interatomic potentials with near first-principles accuracy. These approaches enable large-scale simulations of materials and molecules that were previously computationally infeasible. Our work bridges physics-based modeling and data-driven predictions to unlock new insights in physics, materials science, and chemistry.
AT A GLANCE
- Machine learning models for atomistic simulations
- Near first-principles accuracy with significantly reduced computational cost
- Scalable modeling of complex materials and molecular systems
- Integration of physics-based insights with data-driven approaches
RESEARCH FOCUS
- Developing machine learning interatomic potentials
- Modeling atomic-scale processes in molecules and materials
- Improving accuracy, transferability, and efficiency of ML models
- Developing transferable models across diverse chemical environments
METHODS AND TOOLS
- Machine learning potentials trained on quantum-mechanical reference data
- Equivariant neural networks for atomic interactions
- Large-scale atomistic simulations enabled by ML models
- Computational evaluation against first-principles and experimental results
IMPACT
Machine-learning techniques for atomistic modeling make it possible to study chemical systems at length and time scales far beyond traditional quantum-mechanical simulations. By combining the accuracy of first-principles calculations with the efficiency of data-driven models, these approaches enable deeper insight into atomic-scale mechanisms and accelerate the discovery and design of advanced materials and molecules.
Atomistic modeling in materials science
Quantum-accurate predictions at the cost of classical simulations
We use atomistic modeling to understand and predict how molecules and materials behave at the atomic scale. These simulations complement experiments by revealing mechanisms that are difficult to observe directly, such as defect dynamics, structural transformations, or phase transitions. By combining computational and experimental perspectives, we achieve a deeper and more reliable understanding of material properties and performance.
AT A GLANCE
- Exploring large and complex material landscapes
- Revealing molecular processes with quantum-level fidelity
- Predicting structure–property relationships at scale
- Accelerating discovery across chemistry, physics, and materials science
RESEARCH FOCUS
- Modeling atomic-scale behavior in molecules and solid materials
- Exploring defect motion, phase transitions, and structure evolution
- Connecting microscopic mechanisms with macroscopic material properties
- Supporting experimental insights through complementary simulations
METHODS AND TOOLS
- Atomistic simulations based on established physical models
- Workflow integration with experimental measurements
- Analysis frameworks for interpreting structural and dynamical behavior
IMPACT
Atomistic modeling provides access to mechanisms that are otherwise hidden, enabling more targeted material design and more accurate predictions of performance. This synergy accelerates the development of advanced materials for applications in energy science and electronics.
Data-driven atomistic modeling
Machine learning–enhanced simulations for accelerated materials discovery
A key theme at AMC is new data-driven models that serve to efficiently explore and predict the complex behavior of chemical systems at the atomic scale. By combining machine learning with simulation and experimental data, we uncover structure-property relationships and guide molecular-scale discovery. Using adaptive design of experiments, we strategically select simulations or measurements to maximize information gain and model accuracy. This closed-loop integration of computation and experiment accelerates the development of novel materials with targeted properties.
AT A GLANCE
- Machine learning models for atomistic systems
- Discovering structure–property relationships in chemical materials
- Integration of computational and experimental datasets
- Adaptive strategies to guide simulations and measurements
RESEARCH FOCUS
- Developing machine learning models for molecular and materials data
- Learning structure-property relationships across chemical systems
- Integrating experimental observations with computational predictions
- Designing adaptive workflows that guide data acquisition
METHODS AND TOOLS
- Machine learning models trained on heterogeneous datasets
- Integration of experimental measurements and atomistic simulations
- Data-driven analysis of structural, chemical, and dynamical behavior
IMPACT
Data-driven atomistic modeling enables systematic exploration of complex chemical systems by learning directly from data generated across experiments and computations. This approach reveals fundamental structure–property relationships and enables targeted design of materials with desired performance in areas such as energy, catalysis, and functional materials.
Novel electronic structure methods
Advancing quantum-mechanical modeling beyond conventional density functional theory
We develop novel electronic structure methods that go beyond traditional density functional theory in both accuracy and efficiency. AMC’s vision is designing new approaches to capture complex quantum phenomena – including strong correlations and excited-state dynamicos – while retaining computational efficiency for large-scale simulations. By combining machine learning and linear-scaling techniques, we enable quantum-level insights into systems previously beyond reach. These methods are closely integrated with experimental validation to ensure their impact across materials science, chemistry, and physics.
AT A GLANCE
- Next-generation electronic structure methods beyond standard DFT
- Accurate description of complex quantum phenomena
- Efficient algorithms enabling large-scale quantum computation
- Integration of theory, computation, and experiment
RESEARCH FOCUS
- Developing advanced methods for strongly correlated and excited-state systems
- Improving accuracy and scalability of electronic structure calculations
- Integrating machine learning with quantum-mechanical modeling
- Bridging theoretical predictions with experimental observations
METHODS AND TOOLS
- New quantum-mechanical approaches beyond conventional DFT
- Scalable algorithms for large and complex systems
- Evaluation of theoretical predictions with experimental data
IMPACT
Novel electronic structure methods expand the reach of quantum-mechanical modeling to systems and phenomena that remain inaccessible to conventional approaches. By improving both accuracy and computational efficiency, these developments enable deeper insight into electronic behavior and support the design of advanced materials and molecular systems across chemistry, physics, and materials science.