Improving Machine Learned Force Fields for Complex Fluids through Enhanced Sampling: A Liquid Crystal Case Study

The Journal of Physical Chemistry A, Journal Year: 2024, Volume and Issue: 128(34), P. 7257 - 7268

Published: Aug. 16, 2024

Machine learned force fields offer the potential for faster execution times while retaining accuracy of traditional DFT calculations, making them promising candidates molecular simulations in cases where reliable classical are not available. Some challenges associated with machine include simulation stability over extended periods time and ensuring that statistical dynamical properties underlying simulated systems correctly captured. In this work, we propose a systematic training pipeline such leads to improved model quality, compared achieved by data generation approaches. That relies on use enhanced sampling techniques, it is demonstrated here context liquid crystal, which exemplifies many encountered fluids materials complex free energy landscapes. Our results indicate that, whereas majority field approaches lead dynamics only stable hundred-picosecond trajectories, our approach allows tens nanoseconds organic comprising thousands atoms.

Language: Английский

---

Review of computational advances in tailoring magnesium-hydrogen interactions: Atomistic simulations meet machine learning

International Journal of Hydrogen Energy, Journal Year: 2024, Volume and Issue: 90, P. 114 - 133

Published: Oct. 4, 2024

Language: Английский

---

An Automated Pynta-based Curriculum for ML-Accelerated Calculation of Transition States

Published: Nov. 25, 2024

Microkinetic models (MKMs) are widely used within the computational heterogeneous catalysis community to investigate complex reaction mechanisms, rationalize experimental trends, and accelerate rational design of novel catalysts. However, constructing these requires computationally expensive manually tedious density functional theory (DFT) calculations for identifying transition states each elementary MKM. To address challenges we demonstrate a protocol that uses open-source kinetics workflow tool Pynta automate iterative training reactive machine learning potential (rMLP). Specifically, using silver-catalyzed partial oxidation methanol as prototypical example, first our by an rMLP parallel calculation DFT-quality all 53 reactions, achieving 7x speedup compared DFT-only strategy. Detailed analysis curriculum reveals shortcomings adaptive sampling scheme with single model describe reactions MKM simultaneously. We show limitations can be overcome balanced "reaction class" approach multiple models, describing class similar states. Finally, Pynta-based is also compatible large pre-trained foundational models. For fine-tuning top-performing graph neural network trained on OC20 dataset, observe impressive 20x 89\% success rate in This work highlights synergistic integrating automated tools advance research.

Language: Английский

---

Data efficient machine learning potentials for modeling catalytic reactivity via active learning and enhanced sampling

npj Computational Materials, Journal Year: 2024, Volume and Issue: 10(1)

Published: Dec. 19, 2024

Language: Английский

---

Experimental and Computational Study Toward Identifying Active Sites of Supported SnO_x Nanoparticles for Electrochemical CO₂ Reduction Using Machine‐Learned Interatomic Potentials

Small, Journal Year: 2024, Volume and Issue: unknown

Published: May 25, 2024

SnO

Language: Английский

---