AutoSciLab: A Self-Driving Laboratory For Interpretable Scientific Discovery

TL;DR

AutoSciLab addresses the bottleneck of human-intuition-driven discovery in high-dimensional scientific spaces by integrating a VAE-based generative model, an active-learning loop with Gaussian processes, a directional autoencoder to embed prior physics into a low-dimensional latent space, and a neural-network-based equation learner to express results as human-readable equations. The framework demonstrates that autonomous exploration can rediscover classical physics (projectile motion and Ising magnetization) and uncover novel principles in nanophotonics, while providing interpretable mappings between latent variables and physical quantities. It reports substantial reductions in experimental effort, increased diversity of candidate experiments, and actionable, symbolic insights that tie latent variables to measurable outcomes, positioning AutoSciLab as a scalable path toward automated, interpretable scientific discovery in noisy, expensive domains. The work also discusses limitations, including dependence on curated priors and training data, and outlines a generalizable workflow poised for broad application in physical sciences and engineering.

Abstract

Advances in robotic control and sensing have propelled the rise of automated scientific laboratories capable of high-throughput experiments. However, automated scientific laboratories are currently limited by human intuition in their ability to efficiently design and interpret experiments in high-dimensional spaces, throttling scientific discovery. We present AutoSciLab, a machine learning framework for driving autonomous scientific experiments, forming a surrogate researcher purposed for scientific discovery in high-dimensional spaces. AutoSciLab autonomously follows the scientific method in four steps: (i) generating high-dimensional experiments (x \in R^D) using a variational autoencoder (ii) selecting optimal experiments by forming hypotheses using active learning (iii) distilling the experimental results to discover relevant low-dimensional latent variables (z \in R^d, with d << D) with a 'directional autoencoder' and (iv) learning a human interpretable equation connecting the discovered latent variables with a quantity of interest (y = f(z)), using a neural network equation learner. We validate the generalizability of AutoSciLab by rediscovering a) the principles of projectile motion and b) the phase transitions within the spin-states of the Ising model (NP-hard problem). Applying our framework to an open-ended nanophotonics challenge, AutoSciLab uncovers a fundamentally novel method for directing incoherent light emission that surpasses the current state-of-the-art (Iyer et al. 2023b, 2020).

Figures (11)

• Figure 1: AutoSciLab. Automated 'experiments' are driven by an AL agent sampling the latent space of a generative model (variational autoencoder, VAE) (yellow/purple bubble). 'Experiment' here can refer to a physical laboratory measurement, or a model/simulation of a process. The set of experiments run by the AL agent are distilled using a directional autoencoder to discover a relevant latent space of interest (green bubble). The symbolic relationship between the relevant latent space variables and the physical property of interest is learnt using a neural network equation learner that uses pruning based on connection strength.
• Figure 2: Rediscovering projectile motion. (a) Projectile height $y$ as a function of time $t$. (b) Active learning efficiently finds points with acceleration $\sim$ 10 m/s2. (c) Correlation between initial velocity $u$ and latent space variable $z$ learnt by the directional autoencoder, for trajectories identified by the active learning to have a constant acceleration $g$. (d) Maximum height (H) vs $z$, learnt equation overlaid.
• Figure 3: Rediscovering the spin-dynamics of the Ising spin system. a) Spin-state (s) represented on a grid showing the effects of increasing the temperature. b) Active learning at a fixed temperature ($T \propto (\beta J)^{-1}$) c) $M$ vs $T$, showing overlap between the true equation (black) and the learnt equation (red) from the active learning results (blue).
• Figure 4: Discovering novel relationships in the nanophotonics domain. (a) Generative capacity of the VAE, quantified as the normalized distribution (log scale) of the local slope ($b$) in pump patterns. (b) Directivity as a function of experimental iteration. Dots represent each experiment, and curves reperesent moving averages. (c) Spearman correlations between the discovered latent space and physically relevant pump pattern characteristics. (d) Correlating the latent space with Directivity, incorporating prior knowledge.
• Figure 5: Nanophotonics experimental setup: a) The ultrafast two-color pump-photoluminescence steering setup b) scanning electron microscope image of the nano-fabricated metasurface c) The reflection (blue) and photoluminescence (orange) spectra measured for the fabricated metasurface.