Active Learning in Scientific Discovery: Letting AI Choose the Next Experiment

Introduction

Traditional scientific experiments follow a simple loop: hypothesize, test, analyze, repeat. But what if AI could intelligently choose which experiments to run next, systematically exploring the space of possibilities to reach conclusions faster? This is active learning—and it's transforming how science is done.

The Experimental Bottleneck

Modern science faces a paradox: we can measure more than ever, but we can't measure everything.

The Scale Problem

Consider drug discovery:

• Chemical space: ~10⁶⁰ possible drug-like molecules
• Current libraries: ~10⁸ compounds
• What we can test: ~10³-10⁴ per project

The question: Which tiny fraction should we test?

Traditional approaches:

• Random sampling: Inefficient, misses regions of interest
• Human intuition: Biased, doesn't scale
• Grid search: Combinatorially explosive

Active Learning: The Core Idea

Instead of designing all experiments upfront, active learning operates in cycles:

1. Train model on existing data (however limited)
2. Predict outcomes for all untested possibilities
3. Quantify uncertainty in those predictions
4. Choose next experiment that maximizes information gain
5. Run experiment, add result to dataset
6. Repeat until goal achieved

Key insight: Experiments should be chosen to reduce uncertainty, not just to maximize predicted performance.

Acquisition Functions: Choosing What to Test

The heart of active learning is the acquisition function—a mathematical rule for selecting the next experiment.

1. Uncertainty Sampling

"Test what we're most uncertain about"

def uncertainty_sampling(predictions, uncertainties):
    return argmax(uncertainties)

Pros: Simple, works well for classification Cons: Can waste experiments in irrelevant regions

2. Expected Improvement (EI)

"Test what's most likely to beat our current best"

Used in drug discovery to find molecules better than the current lead compound.

$$EI(x) = E[\max(f(x) - f^*, 0)]$$

Where:

• f(x) = predicted property at x
• f* = best observed value so far

Pros: Balances exploration and exploitation Cons: Can be overly greedy

3. Upper Confidence Bound (UCB)

"Test what might be the best, accounting for uncertainty"

$$UCB(x) = \mu(x) + \beta\sigma(x)$$

Where:

• μ(x) = predicted mean
• σ(x) = predicted uncertainty
• β = exploration parameter

Tuning β:

• High β: More exploration (broader search)
• Low β: More exploitation (local refinement)

4. Thompson Sampling

"Sample according to probability of being optimal"

Randomly samples from the model's posterior distribution.

Advantage: Naturally handles exploration/exploitation trade-off Application: Materials discovery with multiple competing objectives

Bayesian Optimization: The Framework

Bayesian optimization provides the mathematical foundation for active learning in continuous spaces.

Components

Surrogate Model: Predicts outcome and uncertainty

• Gaussian Processes (most common)
• Random Forests
• Neural networks with uncertainty quantification

Acquisition Function: Decides what to test next

Optimization Loop: Iteratively improves both model and experiments

Example: Catalyst Discovery

Chemical reactions require catalysts. Finding optimal catalyst compositions is expensive—each test might cost $10,000 and take days.

Traditional approach: Test 1000 catalysts randomly

• Cost: $10M, 3 years
• Result: 75% yield (best found)

Active learning approach:

• Start with 50 random catalysts
• Use Bayesian optimization to select next 150 tests
• Cost: $2M, 4 months
• Result: 82% yield

Savings: $8M, 2.7 years

Multi-Objective Optimization

Real experiments often optimize multiple objectives:

• Drug discovery: Potency AND safety AND solubility
• Materials: Strength AND conductivity AND cost
• Catalysts: Yield AND selectivity AND stability

Pareto Front

Not a single "best" solution, but a trade-off curve.

qEHVI (Expected Hypervolume Improvement):

• Acquisition function for multi-objective problems
• Efficiently finds Pareto-optimal solutions
• Used at companies like Google and Microsoft

Case Study: Antibiotic Discovery

MIT researchers used multi-objective active learning to discover halicin, a novel antibiotic:

• Objectives: Kill bacteria (efficacy) + Low human toxicity + Novel mechanism
• Search space: 6,000 compounds
• Active learning cycles: 90
• Result: Discovered drug effective against drug-resistant bacteria

Traditional screening: Would have required testing all 6,000 compounds

Constrained Optimization

Sometimes experiments have hard constraints:

• Safety: Don't test explosive/toxic combinations
• Feasibility: Materials must be manufacturable
• Legality: Respect ethical/regulatory limits

Constrained Bayesian Optimization handles these by:

1. Learning constraint functions alongside objectives
2. Restricting search to feasible regions
3. Balancing constraint satisfaction with optimization

Application: Protein Engineering

Optimize enzyme activity subject to:

• Protein must be stable (Tm > 60°C)
• Must be producible in E. coli
• Sequence must be less than 1000 amino acids

Active learning navigates these constraints to find improvements classical methods miss.

Batch Active Learning

Sequential experiments are slow. Can we select batches of experiments to run in parallel?

Challenge: Diversity vs. Quality

If you select 10 experiments purely by uncertainty, they might all be similar (redundant information).

Solution: Diversity-aware batch selection

Techniques

Greedy Sequential Selection

• Pick best experiment
• Update model assuming we already ran it
• Pick next best
• Repeat

Local Penalization

• Reduce acquisition function near selected points
• Encourages diverse sampling

Determinantal Point Processes (DPPs)

• Probabilistic model that favors diversity
• Used in high-throughput screening

Example: Fluorescent Protein Engineering

Optimize brightness in a library of 10⁶ variants:

• Budget: 10,000 experiments
• Lab capacity: 96-well plates (run 96 at once)

Strategy:

• Use active learning to select 100 batches of 96
• Each batch is diverse (covers different regions)
• Result: Found variants 10× brighter than starting protein

Transfer Learning in Active Learning

Experiments in similar domains provide useful prior knowledge.

Scenario: Drug Discovery Campaign

You've optimized molecules for target A. Now you need to optimize for target B (similar but not identical).

Transfer active learning:

1. Pre-train model on target A data
2. Fine-tune on target B experiments
3. Use uncertainty from both to guide selection

Benefit: Requires 50-70% fewer experiments on target B.

Meta-Learning for Acquisition Functions

Can we learn which acquisition function works best for a given problem?

Recent work: Meta-learning algorithms that adapt acquisition functions based on:

• Problem characteristics
• Previous optimization trajectories
• Available budget

Result: 20-30% improvement over fixed acquisition functions.

Self-Driving Labs: Closing the Loop

The ultimate implementation of active learning is the self-driving laboratory—full automation of the experimental cycle.

Components

Robotic Automation: Physical execution of experiments

Real-Time Analysis: Automated measurement and data processing

Active Learning: AI decides next experiments

Closed Loop: No human intervention required

Example: The Organic Synthesis Robot (MIT)

Optimizes chemical reaction conditions:

• Variables: Temperature, concentration, reagent ratios, reaction time
• Search space: 10¹² possible combinations
• Active learning: Bayesian optimization
• Result: Optimized 12 reactions in 24 hours (1/10th the time of manual optimization)

A-Lab (Berkeley)

Autonomously synthesized 41 novel materials:

• Candidates: Selected by active learning from 58,000 possibilities
• Success rate: 71% (vs. 15% for random selection)
• Speed: 17 days (would have taken months manually)

Challenges and Limitations

Model Misspecification

Active learning assumes the model is roughly correct. If it's fundamentally wrong, it leads experiments astray.

Mitigation:

• Include random exploration (ε-greedy strategies)
• Use ensemble models to detect when all models disagree
• Human-in-the-loop checkpoints

Expensive Failures

Some experiments are not just uninformative—they're wasteful or dangerous.

Solution: Risk-aware acquisition functions that penalize high-variance regions.

Non-Stationary Problems

Scientific understanding evolves. Models trained yesterday may not apply today.

Approach: Continual learning with concept drift detection.

The Future: AI-Designed Experiments

Next-generation active learning will:

1. Design Novel Experiments

Not just select from predefined options, but invent new experimental protocols.

2. Integrate Diverse Data

Combine literature, high-throughput screens, expert knowledge, and simulation.

3. Multi-Lab Collaboration

Federated active learning across institutions, sharing insights without sharing data.

4. Causal Discovery

Not just predict outcomes, but infer causal mechanisms.

Practical Implementation

Getting Started

For researchers:

1. Start small: Even 5-10 active learning cycles can outperform random sampling
2. Use existing tools: BoTorch (Facebook), Ax (Meta), GPyOpt
3. Establish baselines: Compare against random and expert-designed experiments
4. Embrace uncertainty quantification: It's not optional—it's essential

Software Tools

BoTorch (PyTorch-based):

from botorch import fit_gpytorch_model
from botorch.acquisition import UpperConfidenceBound
 
# Fit model
model = fit_gpytorch_model(train_X, train_Y)
 
# Select next experiment
UCB = UpperConfidenceBound(model, beta=0.1)
next_x = optimize_acqf(UCB, bounds=bounds)

Summit: For chemical reaction optimization Dragonfly: For high-dimensional problems COMBO: For discrete/combinatorial spaces

Conclusion

Active learning represents a paradigm shift in experimental science: from hypothesis-driven to information-driven research. By systematically quantifying and reducing uncertainty, AI can guide experiments to reach conclusions orders of magnitude faster than traditional approaches.

As self-driving labs become more common and active learning algorithms more sophisticated, we're entering an era where the limiting factor in discovery is not how fast we can run experiments, but how fast we can formulate questions to ask.

The future of science is not just automated—it's intelligently automated, with AI as an equal partner in the scientific process, jointly designing and conducting experiments that push the boundaries of human knowledge.

This article explores how active learning and Bayesian optimization are revolutionizing experimental science, enabling discoveries that would be impossible through traditional trial-and-error approaches—science operating at digital speed.