Machine Learning Framework Emulating Density Functional Theory
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for emulating density functional theory (DFT) using machine learning (ML) struggle to accurately predict both electronic structure and atomic properties, especially for larger systems, and lack a comprehensive scheme for simultaneous prediction of charge density and other DFT properties.
Innovation Solution
A method involving a deep learning scheme that treats the Kohn-Sham equation as an input-output problem, using atomic fingerprints and electron charge density data to predict DFT properties such as potential energy, atomic forces, stress tensor, density of states, valence band maximum, conduction band minimum, and bandgap.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If grid-based schemes are used to predict charge density, then high accuracy is achieved, but computational cost increases and applicability to large databases is hindered
Solution Approach 1:
The patent uses atom-centered basis functions as a simplified copy or representation of the continuous charge density field. Instead of representing charge density on a fine grid for all atoms, the method approximates it using localized atomic orbitals centered on each atom, significantly reducing the number of parameters while maintaining essential physical information.
Solution Approach 2:
The patent applies local quality by using atom-centered basis functions that are localized around each atomic position. This allows the charge density to be represented efficiently with different levels of detail in different regions - high detail near atomic cores where charge density varies rapidly, and coarser representation in interstitial regions, optimizing both accuracy and computational cost.
2Productivity
If atom-based representations are used to predict charge density, then computational cost is reduced and transferability to larger systems is improved, but accuracy decreases especially for delocalized electron densities
Solution Approach 1:
The patent employs a composite representation of charge density that combines atom-centered basis functions with system-wide constraints. The charge density is constructed as a sum of atomic contributions (like a composite material), but with optimization that ensures global consistency and proper description of delocalized electrons, merging local efficiency with global accuracy.
Solution Approach 2:
The patent develops a unified machine learning framework that simultaneously predicts multiple DFT properties (charge density, energy, forces, stress tensor, density of states) using a single model architecture. This multi-functional approach allows the atom-based representation to work effectively across different property predictions and system types, improving both transferability and accuracy.
3Reliability
If traditional DFT methods are used to calculate properties of complex materials with thousands of atoms, then complete electronic structure information is obtained, but the calculations remain inaccessible due to high computational cost
Solution Approach 1:
The patent extracts the essential electronic structure information needed for property prediction from the full DFT formalism. Instead of solving the complete Kohn-Sham equations for large systems, the method extracts key features (charge density distribution, atomic environments) and uses them as inputs to machine learning models that predict target properties, eliminating the most computationally expensive steps while retaining necessary physical information.
Solution Approach 2:
The patent introduces machine learning models as intermediary components between atomic structure and material properties. These ML models act as mediators that take simple atomic descriptors as input and directly output target properties, bypassing the need for explicit electronic structure calculations while maintaining accuracy through training on DFT data.
Data Source
AI summary
A method comprising providing a training data set to a machine learning (ML) system, the training data set indicative of density functional theory (DFT) data for a plurality of materials, the DFT data representing atomic configurations for the plurality of materials, determining a fingerprint for the atomic configuration for each of the plurality of materials, inputting the fingerprints into a machine learning system, generating, with the machine learning system, electron charge density data for the plurality of materials, inputting the electron charge density data and the fingerprint into a machine learning system, and predicting, with the machine learning system, one or more DFT properties of the plurality of materials.