Cause of under performance detection using Bayesian inference in high-dimensions
Basita Das a
a MIT - Massachusetts Institute of Technology, Massachusetts Avenue, 77, Cambridge, United States
Proceedings of Device Physics Characterization and Interpretation in Perovskite and Organic Materials (DEPERO)
València, Spain, 2023 October 3rd - 5th
Organizers: Sandheep Ravishankar, Juan Bisquert and Evelyne Knapp
Invited Speaker, Basita Das, presentation 033
Publication date: 14th September 2023

Bayesian inference methods are useful tools to distinguish between a high number of correlated parameters in a system. They have been previously used in the field of solar cells for well-studied technologies like silicon solar cells, which are already well studied such that a lot of material parameters are well known. It has also been used for material systems like SnS solar cells where the number of unknown parameters is small [1–3]. However, previous implementations were limited by the following challenges - (1) not easily scalable to higher dimensions, (2) solution depending on the initialization condition of the sampler and (3) the speed of operation was also limited by the device model.

In this presentation, we will address these challenges and discuss the strategies we have implemented to overcome them. In our implementation of Bayesian inference, we solve the issue of scalability by using the state-of-the-art Markov chain Monte Carlo (MCMC) sampling technique. However, even though MCMC makes Bayesian inference scalable to higher dimensions, the solution is highly sensitive to the initialization condition of the samplers. To solve this problem, we introduce a hybrid MCMC method coupled with optimization algorithms such that we can maintain robustness of initialization condition from one run to another. This method of initialization is also robust at finding multiple minima/maxima in the solution space. We also incorporated a neural network based surrogate model to replace the device model and hence not limited by the speed of the device model. Overall, as an effect of the three improvements, we achieve improvement in scalability, robustness, and speed even in 15 dimensional problems. The improvements discussed have an overarching impact on high dimensional parameter estimation even in needle-in-haystack like situations.

[1]        R. E. Brandt, R. C. Kurchin, V. Steinmann, D. Kitchaev, C. Roat, S. Levcenco, G. Ceder, T. Unold, and T. Buonassisi, Joule 1, 843 (2017).

[2]        R. Kurchin, G. Romano, and T. Buonassisi, Comput. Phys. Commun. 239, 161 (2019).

[3]        R. C. Kurchin, J. R. Poindexter, V. Vähänissi, H. Savin, ¶ Carlos Del Cañizo, and T. Buonassisi, How Much Physics Is in a Current-Voltage Curve? Inferring Defect Properties from Photovoltaic Device Measurements (2019).

We use our own and third party cookies for analysing and measuring usage of our website to improve our services. If you continue browsing, we consider accepting its use. You can check our Cookies Policy in which you will also find how to configure your web browser for the use of cookies. More info