Research map / Probabilistic and scientific modeling / Molecules and proteins
AutoSciLab: A Self-Driving Laboratory for Interpretable Scientific Discovery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Presents AutoSciLab, a self-driving lab framework chaining a VAE, active learning, a directional autoencoder, and a neural equation learner to rediscover projectile, Ising, and find a new incoherent-light steering equation for metasurfaces.
Problem
Human intuition limits automated labs in designing and interpreting experiments in high-dimensional spaces, so interpretable scientific discovery is bottlenecked.
Approach
Combine a VAE experiment generator, active learning over its latent space, a directional autoencoder that injects physics attributes into latent directions, and a pruned neural equation learner that outputs symbolic equations, following the scientific method end to end.
Key results
- Active learning with EI/UCB reduces the required experiments by roughly an order of magnitude versus differential evolution in the nanophotonics surrogate setting, with discovered pump patterns reaching 3-4x higher directivity than the training set.
- The framework rediscovers the projectile height relation H(u)=u^2/2g and an approximate Ising relation y=1.01 sinh(1.96 beta)+0.1 beta-0.1 close to sinh(2 beta).
- Nanophotonics equation D_e = 0.0467z1^2 - 0.0265z2^2 - 0.175z1 - 0.0955z1z2 + 0.22z2 + 2.707 is reported as a new structure-property relation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use active learning with EI or UCB instead of differential evolution | AutoSciLab nanophotonics search · Nanophotonics | experiments required to find high-directivity patterns: by an order of magnitude | Helped |
| Search the VAE latent space with active learning | AutoSciLab nanophotonics search · Nanophotonics | directivity: 3-4x higher | Helped |
| Generate pump patterns with the VAE rather than using the human-intuition training set | AutoSciLab · Nanophotonics | variety of local slopes: at least two orders of magnitude | Helped |
| Initialize Ising spin states using the active-learning-selected magnetization | AutoSciLab Monte Carlo simulations · Ising Model | Monte Carlo steps to reach equilibrium: order-of-magnitude fewer | Helped |
| Use a directional autoencoder instead of a traditional autoencoder (ablation) | AutoSciLab latent-space interpretation · Nanophotonics | correlation of latent variables with physical quantities | Helped |
| Use a directional autoencoder instead of a traditional autoencoder (ablation) | AutoSciLab latent-space interpretation · Nanophotonics | correlation among latent variables | Helped |
| Use EI rather than UCB as the acquisition function (ablation) | AutoSciLab nanophotonics search · Nanophotonics | breadth of directivity values explored | Helped |
Surprising findings
- The directional autoencoder yields a much stronger latent-physics correlation than a traditional autoencoder (weaker and inter-correlated latents).
- Nanophotonics surrogate-based AL found the optimal sawtooth grating order in about 5 experiments versus brute force over 160.
Limitations
- Human-defined training sets and prior knowledge bound the exploration and may bias results.
- Nanophotonics main results rely on a neural network surrogate rather than physical experiments.
- Gain-factor calculations are approximate and ignore equation-discovery effort.
Experimental setup
Datasets: synthetic projectile trajectories synthetic pump patterns Ising lattice Monte Carlo samples neural surrogate dataset from VAE latent sampling
Benchmarks: projectile motion rediscovery Ising magnetization rediscovery Lotka-Volterra Van der Pol nanophotonics directivity
Metrics: discovered equation accuracy directivity number of experiments gain factor
Baselines: differential evolution upper confidence bound acquisition traditional autoencoder
Builds on: variational autoencoder Gaussian process active learning equation learner differential evolution
Closest papers by meaning
- AutoSciDACT: Automated Scientific Discovery through Contrastive Embedding and Hypothesis Testing (NeurIPS 2025 · similarity 0.86)
- AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation (NeurIPS 2026 · similarity 0.85)
- AI-Researcher: Autonomous Scientific Innovation (NeurIPS 2025 · similarity 0.84)
- End-to-End Phase Field Model Discovery Combining Experimentation, Crowdsourcing, Simulation and Learning (AAAI 2024 · similarity 0.84)
- Adaptive Experimental Design to Accelerate Scientific Discovery and Engineering Design (AAAI 2025 · similarity 0.83)
- LLM and Simulation as Bilevel Optimizers: A New Paradigm to Advance Physical Scientific Discovery (ICML 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.