With advent of Big Data and Machine Learning (ML) methods, there has been much excitement around the application of ML techniques to physical problems. Although ML methods, including deep learning, are powerful for <i>prediction-oriented </i>tasks, we sought to answer if deep-learning methods can “learn” physical laws based on digital data alone. To help answer this question, we used Planck’s black body radiation as a case study, which started the quantum revolution. Multiple neural network architectures were tried to maximize network performance, based on optimizing several hyperparameters including the dimension of the network, activation functions, learning rates, batch size in stochastic gradient descent, etc. Here we report multiple findings related to the feasibility of these machine learning techniques as a proxy for laws of natural sciences. Our analysis indicates the difficulty of extracting physical laws from these approaches and more generally demonstrates a method for studying the functional forms comprising the learned network using the LASSO, thereby helping physically interpret the trained model. Our studies also expose that machine learning is efficient when prior information about the nature of the scientific problem and the context of the correlations are made clear. This in turn provides a systematic way to combine ML with natural sciences for different conservation problems.<br/>

Can machine learning be used to learn laws of natural science? Illustration for Planck’s blackbody radiation

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.