An Integrated Software Package for Studying Structure-Property-Processing Relationships in Material Systems
- Sentient Science Corporation
The identification of structure-property-processing relationships require dynamical models that can access multitude of length spanning nanometers to microns and timescales spanning picoseconds to seconds. Despite its widespread availability of a variety of open source and commercial codes as well as their usage in various flavors, the predictive power of molecular dynamics (MD) is severely limited. Ab-initio MD (AIMD), although very accurate, is extremely challenging to scale beyond a few 100 atoms (< few nm) and tens of picoseconds. Classical MD, on the other hand, is often limited by the accuracy of the inter-atomic potentials and is restricted to nanometer length and nanosecond time scales. Coarse-grained (CG) approaches with reduced degrees of freedom enable us to reach micron/microsecond scales. But the gain in compute efficiency comes at the significant cost of accuracy. Therefore, DOE is seeking an opensource framework that bridge simulations across multiple length and timescales for a given materials system. Sentient Science proposes to develop a scalable open-source high- performance user-friendly web application built upon Argonne’s BLAST – Bridging Length/time scales via Atomistic Simulation Toolkit that enables end users to develop accurate material modes across multiple length/time scales for a wide range of materials. BLAST uses supervised machine learning approach based on genetic algorithms (GA) as well as reinforcement learning algorithms coupled with local optimizers to efficiently (exhaustive yet computationally cheap) scan the parameter landscape to obtain a set of parameters that best reproduce the training dataset. The machine learning algorithms can train potential models or force fields (FF) with training data input from state-of-the-art electronic structure methods (density functional theory) and/or experimental data, coupled with in-house codes and other freely available classical MD packages. A User-friendly workflow/framework for multiscale modeling has been developed built on BLAST. The Sentient team has deployed the framework on AWS micro-services on the DigitalClone platform. The Argonne National Lab’s initial prototype front end is adapted and integrated into DigitalClone platform. The development with focus on making a front end that is suitable for users with diverse computational skills, from novice users with no background in force field development to power users with domain knowledge. The interface simplifies labor-intensive processes in the workflow and address challenges relating to defining the scope of model parameter space, preparing training data, and designing the criterion for quality evaluation. The University of Illinois Chicago (UIC) and Sentient have made significant progress towards the development of reactive force-fields for alloy systems. UIC team has previously demonstrated success of the DC-MDD framework in navigating the potential energy surface of 54 different elements. To enable its application to reactive force-fields of alloys, UIC has added new reactive force-field models and incorporated new multi-objective search algorithms. This software DigitalClone for Materials Design and Discovery (DC-MDD) holds great promise for users within DOE facilities and industrial users in identifying structure-property-processing relationships in various materials classes at an unprecedented pace. The successful implementation of the framework can significantly accelerate the development of new materials at a lower cost and shorter time. Using HPC capacities, Sentient’s DC-MDD software package allows DOE and commercial users with the capability to train and develop their own classical atomistic and coarse-grained interatomic potentials for molecular simulations. This addresses several long-standing problems in the molecular simulation community, such as unintended misuse of existing force fields due to knowledge gap between developers and users, bottlenecks in traditional force field development approaches, and other issues relating to the accuracy, efficiency, and transferability of force fields.
- Research Organization:
- Sentient Science Corporation
- Sponsoring Organization:
- USDOE
- Contributing Organization:
- Argonne National Laboratories
- DOE Contract Number:
- SC0022433
- OSTI ID:
- 1974202
- Type / Phase:
- SBIR (Phase I)
- Report Number(s):
- DOE-SENTIENT-11111
- Country of Publication:
- United States
- Language:
- English
Similar Records
Machine Learning Classical Interatomic Potentials for Molecular Dynamics from First-Principles Training Data
Active Learning A Neural Network Model For Gold Clusters & Bulk From Sparse First Principles Training Data