• Twitter
  • Facebook
  • Google+
  • LinkedIn

Machine-learning-assisted space-transformation accelerates discovery of high thermal conductivity alloys

Name: Dhvaneel Visaria

Department: Mechanical Engineering

Program: B.Tech (4th year)

Name of supervisor: Prof. Ankit Jain

Machine-learning-assisted space-transformation accelerates discovery of high thermal conductivity alloys

Significance

With the advent of superior technology, the need for materials with desired properties (mechanical, electrical, thermal, optical, etc.)  has increased appreciably. The work on materials discovery is relevant and prevalent in all domains from drug discovery and thermoelectricity to electrochemical energy storage and nuclear fuel systems. The research in the field of materials design and discovery has shifted gears from experimental-based work to the extensive utilization of material informatics owing to advancements in computational resources and emergence of powerful data-driven techniques. In this work, we aim to expedite the discovery of high thermal conductivity hypothetical alloys using state-of-the-art generative machine learning technique.

Novelty

There has been much progress in calculating the thermal transport in crystalline solids accurately owing to advancements in the computational resources but the thermal transport in alloys and disordered materials is still unexplored. This is due to the fact that the materials search space for such systems is prohibitively massive in size and therefore, infeasible for exhaustive scanning. Machine learning (ML) is particularly useful in this regard but the traditional surrogate on-the-go ML models still require exponentially large number of model evaluations for these alloys/disordered solids which is not practically feasible. In this study, we demonstrate the applicability of regularized autoencoders, a class of Generative models [as shown in figure alongside], for the exploration of these exponentially large material search spaces.  This model transforms the unit-cell configuration space of the alloy materials to the latent space of the autoencoder in which materials are clustered according to the target property, i.e., thermal conductivity. Leveraging such spatial conditioning of the materials in the latent space, we selectively sample high thermal conductivity materials from the relevant cluster using interpolation.

Methodology

In this study, we focus on a hypothetical system consisting of two-dimensional, graphene-like materials composed of carbon and heavy carbon (twice as heavy as carbon i.e. 24 a.m.u.) atoms in a 32-atom unit cell [as shown in figure alongside]. We aim to discover the unit cell configuration corresponding to very high thermal conductivities. The full thermal conductivity calculations require approximately 10 hours on modern processors and therefore, in practice, these calculations cannot be conducted for all the 232 possible configurations. Therefore, we implement regularized autoencoders, a class of generative models useful in the exploration of these exponentially large material search spaces. This model transforms the unit-cell configuration space of the hypothetical materials to the latent space of the autoencoder in which materials are clustered according to the target property (here – thermal conductivity in x-direction, kx). We leverage such spatial conditioning of the materials in the latent space to selectively sample high thermal conductivity materials from the relevant cluster using interpolation. The iterative materials search process adopted, consists of (i) sampling the materials from the high- kx cluster in the latent space, (ii) validating the kx of these materials using the full thermal conductivity calculations and eventually (iii) re-training the model on the augmented data set until the solution converges to the highest- kx unit cell configuration.

Results/Conclusions

We find that this model performs exceedingly well with an exceptionally low RMSE of 7 W/m-K and well-expedited materials discovery within 10 iterations of the aforementioned search process due to the latent space clustering [as shown in figure]. Also, we observe that the model is able to learn the underlying thermal transport physics of the system under study and is able to predict superlattice-like configurations with high thermal conductivity despite their higher mass. Thus, the proposed methodology can be adopted in search applications for an accelerated and efficient exploration of desired target property.