Machine Learning in Atomistic Computational Material Science Simulation

Original

One may wonder upon looking at this blog’s title, “What do atomistic simulations have to do with machine learning?” For a start, atomistic computational material science (CMS) requires solving the computationally expensive many-body Schrodinger equation. However, this process is done many times on different material structures in a very similar and repetitive manner.

This means that if we can calculate the material properties of ten thousand atomic structures in a computationally expensive manner but use machine learning algorithm to predict the outputs of the remaining million data points, it is a valuable contribution to accelerate the development of materials for different important applications in the society. Machine learning in CMS is basically a machine learning structure-property regression/classification problem!

At Bosch Research, we are interested in predicting functional and reliable materials for energy storage and conversion devices of future electric vehicles such as fuel cells. This means we need to develop simulation tools capable of simulating the movements of ions and chemical reactions on material surfaces.


Atomic motion in molecular dynamics simulation. Atomic forces governing this motion can be quickly predicted by training a neural network on quantum mechanics simulation results.

This kind of simulation (called ab-initio molecular dynamics or AIMD) is very computationally expensive, so we built a neural network capable of replacing quantum mechanics simulation to perform complex atomic structure–force regression in a computationally low-cost manner.

One of the primary challenges in this field is that there are semi-infinite possible input data arrangements, making learning very difficult. Consider the two atomic structures in the figure below. Visually, a human brain immediately understand that they are identical atomic structures; knowing the atomic force vector on the atomic structure on the left also means knowing the atomic force vector for the atomic structure on the left. Mathematically, this means coming up with atomic structure data representation and neural network architecture which exploits the physical properties and symmetries in nature such as translation and rotation symmetries, etc. Algorithms which do not adequately do this will typically suffer from unsatisfactory regression performance.

CUsersmaj5palAppDataLocalMicrosoftWindowsINetCacheContentWordBlog_Picpng

These two different atomic structures are equivalent, but their raw coordinates are different. An ML algorithm which data representation & learning architecture exploit physical properties and symmetries will be more easily trained.

The other challenge is atomic structure feature representation. Unsurprisingly, the field started out with human-assisted feature engineering and quickly (but only very recently) moved into network architectures capable of automated feature extraction, adopting hints and techniques from graph neural networks, image and natural language processing.

This is just one example of the ML-assisted physics-AI hybrid modeling activities done within Bosch Research. “Further research has resulted in products like SoundSee: an algorithm that applies machine learning to hear if something is broken in order to accurately predict machine breakdowns. (Bosch Press 30.01.2019 Link)” AI is one of the core competencies at Bosch Research. In order to underscore its meaning, the Bosch Center for Artificial Intelligence (BCAI) was founded in 2017. Our target is that already in the 2020s, virtually all Bosch products will either be driven by AI themselves, created or designed using AI technology. We work on safe, robust and explainable solutions that contribute to a better quality of life.

****************************************************************

Do you know why Research at Bosch really matters to us?

Research at Bosch really matters to us, because we take an ethical and responsible approach to the task. Our research also really matters, because what is invented within Bosch always aims to improve quality of life. 

Driven by our motivation “Invented for life”, more than 1,700 world-class researchers in 12 locations around the world enable Bosch to find groundbreaking solutions to meaningful problems. In close collaboration with our business divisions and a global network of leading partners, fascinating innovations are created.

In addition to fine expertise in areas such as E-mobility, Quantum Technology and Healthcare Solutions, one of the core competencies of Bosch Research is Artificial Intelligence (AI). In order to underscore its meaning for Bosch, in 2017, the Bosch Center for Artificial Intelligence (BCAI) was founded. Our target is that already in the 2020s, virtually all Bosch products will either be driven by AI themselves, created or designed using AI technology. This results in safe, robust and explainable solutions that contribute to a better quality of life.

Learn more about Bosch Research: www.bosch.com/research or www.bosch-AI.com


Original

Deep Learning Healthcare Deep Learning in Healthcare Summit Medical Imaging


0 Comments

    As Featured In

    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original

    Partners & Attendees

    Intel.001
    Nvidia.001
    Acc1.001
    Rbc research.001
    Mit tech review.001
    Graphcoreai.001
    Twentybn.001
    Forbes.001
    Kd nuggets.001
    Ibm.001
    Facebook.001
    Maluuba 2017.001
    This website uses cookies to ensure you get the best experience. Learn more