Bewertung von atomistischen Simulationen
Zusammenfassung der Projektergebnisse
The primary achievement of the Scientific Network discussions lies in the collective experience of the development, validation, and application of machine learning interatomic potentials (MLIP). The key advancements and crucial steps in this process include: Applicability Domain Identification: A pivotal decision in the MLIP development process is to determine the future applicability domain—whether it's locally-applicable (LA) or general-purpose (GP). Reference Dataset Creation: For LA potentials, a dataset comprising structures that describe specific phases, pressures, and temperature conditions is essential. For the GP potentials, the dataset is substantially larger, encompassing a broader spectrum of phases, compositions, densities, and various defect types. High-precision and consistent DFT calculations are critical for all these structures. The use of automated or autonomous frameworks for high-throughput calculations substantially enhances robustness and reproducibility. MLIP Parameterization: The parameterize the initial version of the MLIP, specific tools, such as PACEmaker, RuNNer, AtomicRex, etc. are employed. A uniform weights distribution over samples in the training set is recommended for LA potentials, whereas a non-uniform weights distribution, with more focus on more relevant structures, suits GP potentials better. Basic Validation: Basic validation of the initial MLIP includes assessing properties directly comparable to reference DFT. Deviations between the predicted properties with MLIP and reference DFT data can be addressed by adding more property-relevant structures to the training set and adjusting weights in the loss function during parameterization. Advanced Validation: Advanced validation entails the use of long-time and/or large-scale molecular dynamics simulations, with parallel uncertainty indication (UI) by the corresponding method for the given MLIP. The structures with large uncertainties form the pool of candidates. From this pool of candidates, the most representative structures can be further chosen using various quantification methods for local atomic environments. These selected structures are recalculated using the reference DFT method, added to the training set and then MLIP is reparameterized. Active Learning Cycle: Iterating the steps above continuously advances the construction of the training set and the development of a robust machine learning interatomic potential. These steps are foundational, adaptable, and readily automatable, reflecting the recent advancements in the field of MLIP. The COVID-19 pandemic disrupted our initial plan of holding in-person meetings and consequently led to a shift in the project timeline. During this period, the scientific landscape in the field of atomistic simulations experienced significant transformations due to the emergence and widespread adoption of cutting-edge machine learning interatomic potentials. As a result, our scientific network meetings refocused their agenda from a general evaluation of existing interatomic potentials to the methodologies for parameterizing machine learning potentials in a broader sense, including dataset generation, potential parameterization, validation, active learning and application.
Projektbezogene Publikationen (Auswahl)
-
Phase transitions in titanium with an analytic bond-order potential. Modelling and Simulation in Materials Science and Engineering, 27(8), 085008.
Ferrari, Alberto; Schröder, Malte; Lysogorskiy, Yury; Rogal, Jutta; Mrovec, Matous & Drautz, Ralf
-
Transferability of interatomic potentials for molybdenum and silicon. Modelling and Simulation in Materials Science and Engineering, 27(2), 025007.
Lysogorskiy, Y.; Hammerschmidt, T.; Janssen, J.; Neugebauer, J. & Drautz, R.
-
Optimized interatomic potential for study of structure and phase transitions in Si-Au and Si-Al systems. Computational Materials Science, 184, 109891.
Starikov, S.; Gordeev, I.; Lysogorskiy, Y.; Kolotova, L. & Makarov, S.
-
Efficient parametrization of the atomic cluster expansion. Physical Review Materials, 6(1).
Bochkarev, Anton; Lysogorskiy, Yury; Menon, Sarath; Qamar, Minaam; Mrovec, Matous & Drautz, Ralf
-
Multilayer atomic cluster expansion for semilocal interactions. Physical Review Research, 4(4).
Bochkarev, Anton; Lysogorskiy, Yury; Ortner, Christoph; Csányi, Gábor & Drautz, Ralf
-
Active learning strategies for atomic cluster expansion models. Physical Review Materials, 7(4).
Lysogorskiy, Yury; Bochkarev, Anton; Mrovec, Matous & Drautz, Ralf
