Autonomous Science: Big Data Tools for Small Data Problems in Chemistry

Abstract

Machine learning tools are emerging to support autonomous science, in which critical decision-making on experimental design is conducted by algorithms rather than by human intervention. This shift from automation to autonomation is enabled by rapid advances in data science and deep neural networks, which provide new strategies for mining the ever-increasing volumes of data produced by modern instrumentation. However, a large number of measurements are intrinsically incompatible with high-throughput analyses, limited by time, the availability of materials, or the measurement architecture itself. Counter-intuitively, strategies developed for big-data challenges have the potential for major impacts in such data-limited problems. Two strategies for leveraging “big data” tools for small data challenges form the central theme of this chapter. In the first, advances in autonomous design of experiments are reviewed, in which algorithms select in real-time the next most informative experiments to perform based on results from previous measurements. Autonomous science enables maximization of confidence in scientific decision-making while simultaneously minimizing the number of measurements required to achieve that confidence. In the second, recent advances in adversarial strategies are reviewed for improving chemical decision-making with limited data. Adversarial attacks can help identify weak-points in classification and dimension reduction approaches that naturally arise in data-sparse training. Once identified, generative adversarial approaches provide a framework for “shoring up” those weak points by optimally leveraging the underlying probability distributions describing the input data. These illustrative examples highlight the rapidly evolving landscape of chemical measurement science enabled by machine learning.

On-the-fly closed-loop materials discovery via Bayesian active learning

“…Taking another step and placing active learning in real-time control of solid-state materials exploration labs promises to accelerate materials discovery while also rapidly and efficiently illuminating complex materials-property relationships. Such potential innovation has been discussed in recent prospectives 25 , 26 , with a primary focus on autonomous chemistry 27 – 29 .…”

Functional Group Identification for FTIR Spectra Using Image-Based Machine Learning Models

“…9,[15][16][17] Generalizable functional group ML models would increase the utility of FTIR sample screening in environmental and other chemistry applications. 18,19 In this study, we investigate the implementation of convolutional neural networks (CNNs) 20 to identify functional groups present in FTIR spectra. By limiting spectral preprocessing, we explore a minimalistic approach to allow the network to learn spectral patterns for successful recognition of the fifteen most common organic functional groups (Table 1).…”

“…There is thus an unexplored, yet applicable field of FTIR spectral interpretation through statistical methods. Progress toward machine learning (ML) methods for environmental pollutant analysis has been explored for specific, targeted applications. ,− Generalizable functional group ML models would increase the utility of FTIR sample screening in environmental and other chemistry applications. , …”

Machine Learning in Chemical Product Engineering: The State of the Art and a Guide for Newcomers

“…This lack of data is quite frequent in CPE problems. To overcome this limitation, the scientific community is looking for ML methods that are specifically adapted to limited-data problems, such as kernel methods, low-variance models with feature reduction capabilities, multi-process modeling and transfer learning [189][190][191][192]. An example is given in [10], where a DNN, implemented organic materials design, updates its initial weights from a large data set, derived from a similar domain to the target problem, and then fine tunes its weights using a smaller, dedicated data set, thus, learning the subtle characteristics that are specific to the targeted application.…”