LLM-AutoSciLab: Closed-Loop Scientific Discovery via Active Experimentation with LLMs

Scientific discovery is a closed-loop process in which hypotheses guide data acquisition and observations refine the hypothesis space. Yet most approaches reduce discovery to supervised learning over fixed datasets, where limited observations can support multiple plausible mechanisms that fit locally but fail to generalize. Thus, the key challenge is selecting informative observations to resolve uncertainty, shifting the focus from static inference to adaptive data acquisition. To address this, we propose LLM-AutoSciLab, a closed-loop framework that couples hypothesis generation with hypothesis-conditioned experiment selection and mechanism refinement. Rather than fitting models to passively collected data, LLM-AutoSciLab iteratively proposes plausible hypotheses, selects informative experiments to distinguish or refine them, and updates its state using the resulting evidence. To evaluate dynamic, closed-loop scientific discovery with active data acquisition, we introduce ActiveSciBench, comprising two datasets: ActiveSciBench-Chem with 57 enzyme-kinetics tasks and ActiveSciBench-GRN with 45 gene-regulatory-network tasks. These datasets model discovery as a budget-constrained process requiring adaptive experiment design, variable selection, and recovery of true mechanisms. Across NewtonBench, ActiveSciBench-Chem, and ActiveSciBench-GRN, LLM-AutoSciLab outperforms prior methods, achieving 67.6% and 35.1% symbolic accuracy on NewtonBench and ActiveSciBench-Chem, respectively, and 31.1% exact graph recovery on ActiveSciBench-GRN. Moreover, hypothesis-guided experimentation is 2-5x more sample-efficient than the strongest competing baselines. Code and data are available at: https://github.com/scientific-discovery/LLM-AutoSciLab

Boltzmann Reinforcement Learning for Noise resilience in Analog Ising Machines

Analog Ising machines (AIMs) have emerged as a promising paradigm for combinatorial optimization, utilizing physical dynamics to solve Ising problems with high energy efficiency. However, the performance of traditional optimization and sampling algorithms on these platforms is often limited by inherent measurement noise. We introduce BRAIN (Boltzmann Reinforcement for Analog Ising Networks), a distribution learning framework that utilizes variational reinforcement learning to approximate the Boltzmann distribution. By shifting from state-by-state sampling to aggregating information across multiple noisy measurements, BRAIN is resilient to Gaussian noise characteristic of AIMs. We evaluate BRAIN across diverse combinatorial topologies, including the Curie-Weiss and 2D nearest-neighbor Ising systems. We find that under realistic 3\% Gaussian measurement noise, BRAIN maintains 98\% ground state fidelity, whereas Markov Chain Monte Carlo (MCMC) methods degrade to 51\% fidelity. Furthermore, BRAIN reaches the MCMC-equivalent solution up to 192x faster under these conditions. BRAIN exhibits $\mathcal{O}(N^{1.55})$ scaling up to 65,536 spins and maintains robustness against severe measurement uncertainty up to 40\%. Beyond ground state optimization, BRAIN accurately captures thermodynamic phase transitions and metastable states, providing a scalable and noise-resilient method for utilizing analog computing architectures in complex optimizations.

Accelerating Materials Design via LLM-Guided Evolutionary Search

Materials discovery requires navigating vast chemical and structural spaces while satisfying multiple, often conflicting, objectives. We present LLM-guided Evolution for MAterials design (LLEMA), a unified framework that couples the scientific knowledge embedded in large language models with chemistry-informed evolutionary rules and memory-based refinement. At each iteration, an LLM proposes crystallographically specified candidates under explicit property constraints; a surrogate-augmented oracle estimates physicochemical properties; and a multi-objective scorer updates success/failure memories to guide subsequent generations. Evaluated on 14 realistic tasks spanning electronics, energy, coatings, optics, and aerospace, LLEMA discovers candidates that are chemically plausible, thermodynamically stable, and property-aligned, achieving higher hit-rates and stronger Pareto fronts than generative and LLM-only baselines. Ablation studies confirm the importance of rule-guided generation, memory-based refinement, and surrogate prediction. By enforcing synthesizability and multi-objective trade-offs, LLEMA delivers a principled pathway to accelerate practical materials discovery. Code: https://github.com/scientific-discovery/LLEMA

AutoSciLab: A Self-Driving Laboratory For Interpretable Scientific Discovery

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).

Self-driving lab discovers principles for steering spontaneous emission

We developed an autonomous experimentation platform to accelerate interpretable scientific discovery in ultrafast nanophotonics, targeting a novel method to steer spontaneous emission from reconfigurable semiconductor metasurfaces. Controlling spontaneous emission is crucial for clean-energy solutions in illumination, thermal radiation engineering, and remote sensing. Despite the potential of reconfigurable semiconductor metasurfaces with embedded sources for spatiotemporal control, achieving arbitrary far-field control remains challenging. Here, we present a self-driving lab (SDL) platform that addresses this challenge by discovering the governing equations for predicting the far-field emission profile from light-emitting metasurfaces. We discover that both the spatial gradient (grating-like) and the curvature (lens-like) of the local refractive index are key factors in steering spontaneous emission. The SDL employs a machine-learning framework comprising: (1) a variational autoencoder for generating complex spatial refractive index profiles, (2) an active learning agent for guiding experiments with real-time closed-loop feedback, and (3) a neural network-based equation learner to uncover structure-property relationships. The SDL demonstrated a four-fold enhancement in peak emission directivity (up to 77%) over a 72{\deg} field of view within ~300 experiments. Our findings reveal that combinations of positive gratings and lenses are as effective as negative lenses and gratings for all emission angles, offering a novel strategy for controlling spontaneous emission beyond conventional Fourier optics.

Rethinking materials simulations: Blending direct numerical simulations with neural operators

Direct numerical simulations (DNS) are accurate but computationally expensive for predicting materials evolution across timescales, due to the complexity of the underlying evolution equations, the nature of multiscale spatio-temporal interactions, and the need to reach long-time integration. We develop a new method that blends numerical solvers with neural operators to accelerate such simulations. This methodology is based on the integration of a community numerical solver with a U-Net neural operator, enhanced by a temporal-conditioning mechanism that enables accurate extrapolation and efficient time-to-solution predictions of the dynamics. We demonstrate the effectiveness of this framework on simulations of microstructure evolution during physical vapor deposition modeled via the phase-field method. Such simulations exhibit high spatial gradients due to the co-evolution of different material phases with simultaneous slow and fast materials dynamics. We establish accurate extrapolation of the coupled solver with up to 16.5$\times$ speed-up compared to DNS. This methodology is generalizable to a broad range of evolutionary models, from solid mechanics, to fluid dynamics, geophysics, climate, and more.

Parsimonious neural networks learn classical mechanics and can teach it

We combine neural networks with genetic algorithms to find parsimonious models that describe the time evolution of a point particle subjected to an external potential. The genetic algorithm is designed to find the simplest, most interpretable network compatible with the training data. The parsimonious neural network (PNN) can numerically integrate classical equations of motion with negligible energy drifts and good time reversibility, significantly outperforming a generic feed-forward neural network. Our PNN is immediately interpretable as the position Verlet algorithm, a non-trivial integrator whose justification originates from Trotter's theorem.