Aceleración del Descubrimiento de Materiales Mediante Potenciales de Aprendizaje Automático con Procesos Gaussianos Dispersos | JoVE Visualize

Área de la Ciencia:

Computational Materials Science
Machine Learning in Chemistry
Quantum Mechanics

Sus antecedentes:

Quantum mechanical calculations provide high accuracy but are computationally expensive, limiting simulations to small systems.
Materials discovery for advanced applications like batteries and solar cells requires simulations at realistic scales.
Existing machine learning potentials often need extensive training data, posing a bottleneck for widespread adoption.

Objetivo del estudio:

To present sparse Gaussian process regression (SGPR) as a statistically rigorous machine learning framework for materials simulations.
To achieve quantum-level accuracy using minimal training data and provide calibrated uncertainty estimates.
To enable faster and more accurate simulations for materials discovery.

Principales métodos:

Developed a sparse Gaussian process regression (SGPR) framework utilizing rank reduction and intelligent selection of informative chemical environments.
Implemented an on-the-fly adaptive sampling strategy to trigger new quantum calculations based on model uncertainty.
Employed a robust Bayesian committee machine (RBCM) architecture to partition and combine specialized expert models for complex systems.

Principales resultados:

SGPR achieved practical accuracy with 100-1000 quantum calculations, significantly reducing data requirements compared to other methods.
Demonstrated SGPR's versatility in simulating solid electrolytes (Li7P3S11), perovskite solar cells, electrocatalysts (Pt-C2N2), and organic systems.
Achieved substantial speedups (up to 10^4x) and revealed mechanistic insights, such as stabilizing interlayers in perovskite solar cells.

Conclusiones:

The SGPR-RBCM framework offers significant advantages for materials simulations when training data is limited and uncertainty quantification is crucial.
Enables quantum-accurate simulations at near-classical computational costs, accelerating high-throughput screening.
Provides a pathway toward comprehensive machine learning potentials for transforming materials discovery in clean energy and electronics.