Is machine learning good or bad for the natural sciences?

TL;DR

The paper examines how the data-centric, predictive paradigm of machine learning aligns with the natural sciences' goal of understanding latent structure. It identifies two hard-to-avoid biases—training-set bias amplification and emulator-induced confirmation bias—and argues for careful, domain-aware application of ML, especially in causal modeling and instrument calibration. It outlines safe, high-value uses such as real-time decision making, outlier discovery, and nuisance modeling, while warning against overreliance on emulators and label-transfer approaches for population inferences. The authors advocate a physics-informed, interpretable, and reproducible approach to ML in science, concluding that ML is both essential and potentially perilous depending on the context.

Abstract

Machine learning (ML) methods are having a huge impact across all of the sciences. However, ML has a strong ontology - in which only the data exist - and a strong epistemology - in which a model is considered good if it performs well on held-out training data. These philosophies are in strong conflict with both standard practices and key philosophies in the natural sciences. Here we identify some locations for ML in the natural sciences at which the ontology and epistemology are valuable. For example, when an expressive machine learning model is used in a causal inference to represent the effects of confounders, such as foregrounds, backgrounds, or instrument calibration parameters, the model capacity and loose philosophy of ML can make the results more trustworthy. We also show that there are contexts in which the introduction of ML introduces strong, unwanted statistical biases. For one, when ML models are used to emulate physical (or first-principles) simulations, they amplify confirmation biases. For another, when expressive regressions are used to label datasets, those labels cannot be used in downstream joint or ensemble analyses without taking on uncontrolled biases. The question in the title is being asked of all of the natural sciences; that is, we are calling on the scientific communities to take a step back and consider the role and value of ML in their fields; the (partial) answers we give here come from the particular perspective of physics.

Figures (1)

• Figure 1: Visualization of the toy regression. Top-left: Random examples of the data vectors $x$, which are one-dimensional images generated from a linear model plus a nonlinearity created by two rectifications. Details of the data generation are given in the text. Each example $x$ is labeled on the right side by the value of its label $y$. Top-right: The training-set labels $y$, plotted against the known parameter $r$, which is not used in the regression (only the vectors $x$ are used). Also shown are solid red circles showing the true mean relationship between $y$ and $r$ in the toy data. Open black squares show the empirical mean relationship measured in bins in the training-set data. Bottom-left: Validation of the trained regression in the validation set, showing that the label estimates $\hat{y}$ are noisy (as expected given the problem set-up) but not strongly biased. Bottom-right: The regression estimates $\hat{y}$ in the very large test set, plotted against the known parameter $r$. Also shown are the same solid red circles and open black squares as in the top-right plot. Black X-shaped symbols show the mean relationship between $\hat{y}$ and $r$. The relationship shown by the Xs is very precise (error bars are much smaller than the symbols; see the text) but biased far away from the true relationship, unlike the relationship shown by the open squares (measured in the training set alone).