Bayesian Optimization for Autonomous Experiments (2026)

A self-driving lab can run a synthesis-and-measure cycle every few hours, but the physical world does not care how fast your robots move. Each experiment burns reagents, machine time, and often a rare precursor you cannot re-order. The scarce resource is not compute — it is the experiment itself. Bayesian optimization for autonomous experiments is the planning brain that decides which single point in a vast design space is worth spending the next cycle on, given everything measured so far. It is the difference between a lab that blunders through a grid search for a year and one that converges on a strong candidate in forty runs.

This matters now because autonomous science has moved from demo to deployment. Robotic platforms for materials, catalysis, and formulation chemistry are online, and the bottleneck has shifted from can we automate the pipette to are we choosing good experiments. A fast robot running a dumb planner just wastes reagents faster. The planner is where the leverage sits.

What this covers: the surrogate model, acquisition functions, the explore-exploit trade-off, batch and cost-aware optimization, constraints and categorical variables, a worked acquisition decision, a BoTorch-style snippet, and the failure modes that quietly wreck real campaigns.

Context and Background

Classical design of experiments — full factorial grids, Latin hypercubes, response-surface methods — assumes experiments are cheap enough to tile the space evenly. That assumption breaks the moment a single run costs hundreds of dollars and half a day. If your design space has six continuous knobs and you allow ten levels each, a grid is a million experiments. No lab runs a million experiments. You get tens, maybe a few hundred.

Bayesian optimization (BO) is the response to that scarcity. It is a sequential, model-based strategy for optimizing an expensive black-box function: you have no gradient, no closed form, and every evaluation is precious. BO builds a probabilistic surrogate of the objective from the handful of points measured so far, then uses that surrogate to reason about where the next measurement will most improve your knowledge or your best result. It has a long lineage — Kushner’s work in the 1960s, Mockus in the 1970s, and the machine-learning revival after 2010 when it became the standard tool for hyperparameter tuning.

The theoretical reason BO works is that it minimizes cumulative regret — the gap between what you measured and the true optimum, summed over the campaign. A well-chosen acquisition function drives that regret down at a rate that grid search and random search cannot match, because those methods ignore everything they have already learned. BO instead compounds its knowledge: every measurement sharpens the surrogate, which sharpens the next decision. On a smooth six-dimensional response, a good BO campaign routinely finds a strong optimum in tens of experiments where a grid would need thousands. That efficiency is not a marginal gain; it is the difference between a feasible campaign and an impossible one.

The autonomous-science community adopted BO because the fit is almost perfect. A self-driving lab is a closed loop that proposes, executes, and measures; BO supplies the propose step. The broader loop — scheduling, robotic execution, data capture — is covered in our self-driving lab architecture guide. Open-source tooling matured alongside: BoTorch and Ax from Meta, scikit-optimize, GPyOpt, and Dragonfly for general use, plus chemistry-aware tools like Gryffin and Atlas from the Aspuru-Guzik group. The Olympus package even provides benchmark surfaces so competing planners can be compared on realistic, noisy response landscapes rather than toy functions. That ecosystem is why a small lab can stand up a credible planner in an afternoon.

It helps to place BO against its neighbours. Reinforcement learning also learns from interaction, but it is built for problems with thousands or millions of cheap steps and a sequential state that evolves — the opposite of a lab where each step is expensive and largely independent. Evolutionary and genetic algorithms optimize black boxes too, but they are sample-hungry, needing large populations across many generations, which is fine in simulation and ruinous when each evaluation is a physical synthesis. BO’s defining assumption is the expensive evaluation: it invests heavily in modelling to squeeze the maximum decision quality out of every single measurement. That is exactly the regime a self-driving lab lives in, which is why BO, not RL or evolutionary search, became the planner of choice for autonomous experimentation.

The Bayesian Optimization Loop as an Experiment Planner

Bayesian optimization is a two-part machine: a surrogate model that predicts the objective and its uncertainty everywhere in the design space, and an acquisition function that turns those predictions into a single number scoring how useful each candidate experiment would be. The planner fits the surrogate to all data so far, maximizes the acquisition function to pick the next point, runs that experiment, adds the result, and repeats. That closed loop is the whole idea.

Figure 1: The BO planner sits inside the self-driving-lab loop, converting measured data into the next experiment.

Figure 1 shows where the planner lives. The design space and its constraints feed a surrogate; the surrogate feeds an acquisition function; the acquisition function picks one experiment; the robotic platform synthesizes and characterizes it; the new labelled measurement updates the dataset; and the loop closes back to the surrogate. The planner never touches a pipette — it only decides. Everything physical happens in the robot, and everything smart happens in these two boxes.

The surrogate: a Gaussian process gives you mean and uncertainty

The default surrogate is a Gaussian process (GP). A GP treats the unknown objective as a distribution over functions, and after conditioning on your observations it returns, for any candidate point, a predicted mean and a predicted standard deviation. That second number is the whole reason GPs dominate BO. An acquisition function cannot balance exploration against exploitation without a calibrated estimate of how unsure the model is at each point.

A GP is defined by its kernel, which encodes your assumption about how smoothly the objective varies. The Matern 5/2 kernel is the workhorse in BO because it assumes twice-differentiable but not infinitely smooth functions — a realistic prior for physical responses. The kernel’s length-scales, fitted by maximizing the marginal likelihood, tell you how far a measurement’s influence reaches across each dimension. Get the length-scale wrong and the whole campaign wanders.

Mechanically, the GP builds a kernel matrix whose entries are the covariance between every pair of observed points. Predicting at a new point requires inverting that matrix, an operation that costs on the order of n-cubed for n observations. For the tens or low hundreds of experiments a physical lab runs, that cost is trivial — inversion takes milliseconds and is dwarfed by the day-long experiment it plans. This is the counterintuitive economics of BO: you spend serious modelling effort per decision precisely because the decision is worth so much. The cubic scaling only becomes a wall in high-throughput virtual screening with thousands of points, where sparse GPs or alternative surrogates take over.

When the GP breaks: high dimensions and categorical spaces

GPs are excellent up to roughly ten to twenty continuous dimensions and small datasets, because their cost scales cubically with the number of observations. Beyond that, or when variables are categorical — which ligand, which solvent, which of eight synthesis routes — a plain GP struggles. Two responses matter here. Random-forest surrogates (the approach popularized by SMAC) handle categorical and conditional variables naturally and scale to more data. Bayesian neural networks and deep-ensemble surrogates give calibrated uncertainty in high dimensions where GPs choke. Chemistry-specific tools such as Gryffin were built precisely for categorical and mixed spaces, using a kernel-density formulation over discrete choices.

The acquisition function converts belief into a decision

Given the surrogate’s mean and standard deviation everywhere, the acquisition function assigns each candidate a scalar score, and the planner runs whichever candidate scores highest. This is the actual acquisition function experiment design step — the moment belief becomes a committed action. The design space might be continuous, so “maximize the acquisition” is itself an inner optimization, usually solved by multi-start gradient ascent on the cheap surrogate rather than on the real experiment. That inner loop costs milliseconds; the outer loop costs a day. That asymmetry is the entire economic argument for BO.

Initialization and stopping are part of the planner

Two design choices bracket the loop and get far too little attention. The first is initialization. A GP fitted to two or three points is almost useless, so you seed the campaign with a space-filling design — Sobol sequences or a Latin hypercube — of roughly five to ten points per dimension. Too few seed points and the first surrogate is fitting noise; too many and you have burned budget on a blind grid before the planner ever engages. The second is stopping. Naive campaigns just run until the budget is exhausted, but a smarter loop watches the acquisition value at the chosen point each round. When the best available Expected Improvement falls below a small threshold, the surrogate is telling you it expects no meaningful gain from any further experiment, and continuing is waste. A convergence-aware stop can return reagents and machine time to other campaigns, which in a shared lab is real money.

Acquisition Functions and the Explore-Exploit Trade-off

Directly answered: an acquisition function scores each candidate experiment by combining the surrogate’s predicted value with its uncertainty, so the planner can pick the point that best trades chasing the current best (exploit) against probing the unknown (explore). Different acquisition functions weight that trade-off differently, and choosing one is the planner’s most consequential design decision.

Figure 2: The surrogate’s mean and standard deviation feed the acquisition score, which resolves the explore-exploit balance into one next point.

Figure 2 traces the mechanism. Observed data fit the surrogate; the surrogate emits a posterior mean and standard deviation at every candidate; both feed the acquisition score; and the explore-exploit balance decides whether the winning point sits in a high-uncertainty region (explore, to shrink uncertainty) or a high-mean region (exploit, to bank a good result). The argmax of the acquisition surface is the next experiment.

Expected Improvement, UCB, and Probability of Improvement

Expected Improvement (EI) is the default. It computes the expected amount by which a candidate will beat the current best, integrating over the surrogate’s predictive distribution. EI has a clean closed form for a GP and naturally balances the two forces: a point scores well either because its mean is high or because its uncertainty is large enough that it might be high. It is the sensible starting choice for most single-objective campaigns.

Upper Confidence Bound (UCB) scores a point as its mean plus a tunable multiple of its standard deviation, written mean plus beta times sigma. Beta is an explicit explore-exploit dial: small beta exploits, large beta explores. UCB is transparent and easy to reason about, which is why practitioners who want a hand on the wheel often prefer it over EI’s implicit balance.

Probability of Improvement (PI) scores the probability that a candidate beats the current best by some margin. PI is the most exploitative of the three and tends to get stuck refining a known peak unless you inflate its margin. It is rarely the best default but is useful late in a campaign when you want to squeeze a local optimum.

Thompson sampling and information-based acquisition

Thompson sampling takes a different route: draw one random function from the surrogate’s posterior and pick the point that maximizes that sample. Because each draw is a plausible version of reality, the method explores in proportion to genuine uncertainty, and it parallelizes beautifully — each parallel worker just draws its own sample. That property makes it a favourite for high-throughput labs.

Information-based acquisitions — entropy search, predictive entropy search, and max-value entropy search — score a candidate by how much it is expected to reduce uncertainty about the location or value of the optimum, measured in bits. They are the most principled choice when a single probe is extraordinarily expensive and you want every measurement to buy maximum information, though they cost more to compute than EI or UCB.

Multi-objective: qEHVI when you optimize two things at once

Real experiments rarely optimize one number. You want high yield and low cost, or high conductivity and mechanical stability. Multi-objective BO seeks the Pareto frontier — the set of solutions where you cannot improve one objective without sacrificing another. The standard acquisition here is qEHVI, q-Expected Hypervolume Improvement, which scores a batch of candidates by how much they are expected to expand the volume dominated by the Pareto frontier. BoTorch’s qEHVI implementation is the common reference, and it handles noisy observations through its qNEHVI variant. This is the right tool whenever a scientist says “optimize A but do not wreck B.”

An older alternative is scalarization: collapse the objectives into one weighted sum and run ordinary single-objective BO. It is simple and cheap, but it has a real weakness — a fixed weighting can only ever find points on the convex parts of the Pareto frontier, and it forces you to commit to a trade-off ratio before you know the shape of the frontier. Random scalarizations, where the weights are resampled each round, partly fix this by sweeping the frontier over time. Hypervolume-based methods like qEHVI avoid the problem entirely by rewarding frontier coverage directly, which is why they have become the default when you can afford the extra computation. The practical rule: reach for scalarization only when objectives are few and their relative importance is genuinely known in advance; otherwise let hypervolume find the frontier for you.

Batch, Parallel, and Cost-Aware Bayesian Optimization

Textbook BO is sequential: propose one, measure one, repeat. A real self-driving lab runs a plate of 24 or 96 wells at once. Proposing a single experiment while 95 robotic channels sit idle wastes the platform. Batch parallel bayesian optimization solves this by selecting q points per round that are individually promising yet jointly diverse, so the batch does not squander capacity on 96 near-identical runs.

Figure 3: Sequential BO proposes one point per fit; batch BO proposes q points per fit, matching the parallelism of the physical platform.

Figure 3 contrasts the two regimes. Sequential BO fits, picks one point, runs one experiment, and loops — correct but throughput-starved when the hardware is parallel. Batch BO fits, selects q points using a batch-aware acquisition, runs all q at once, collects q results, and loops. The wall-clock cost of a campaign collapses when the batch matches the plate size.

How batch selection avoids redundant experiments

The naive approach — take the top-q points by acquisition score — fails, because the top q candidates cluster around the same peak and you learn almost nothing new from 95 of them. Three techniques fix this. q-EI (and q-EHVI for multi-objective) computes the joint expected improvement of a set, so it rewards diversity directly; BoTorch evaluates it by Monte Carlo sampling over the joint posterior. Local penalization picks points greedily but subtracts a penalty around each already-selected point, pushing the next pick away. Thompson batching draws q independent posterior samples and maximizes each, giving natural spread because each sample disagrees slightly about where the optimum lies. All three turn a wasteful cluster into an informative spread.

Cost-aware BO: not every experiment costs the same

BO’s default assumes every evaluation costs the same, which is often false. One synthesis route takes twenty minutes; another needs an overnight anneal. One precursor is cheap; another is a milligram of something precious. Cost-aware BO divides the acquisition score by an estimated cost — EI-per-unit-cost is the canonical form — so the planner prefers cheap, informative experiments and only spends on an expensive one when its expected payoff justifies the price. When cost itself is unknown, you can model it with a second GP and optimize expected improvement per expected cost. This reframes the campaign budget from “number of experiments” to “total resource spent,” which is what a lab manager actually cares about.

Multi-fidelity BO: cheap proxies before costly experiments

A close cousin of cost-aware BO is multi-fidelity optimization, and it matters enormously in real labs. Often you have several ways to evaluate a candidate at different cost and accuracy: a fast physics simulation, a quick low-resolution measurement, or a slow gold-standard characterization. Multi-fidelity BO models all fidelities jointly and lets the planner spend cheap, approximate evaluations to narrow the field, then commit expensive high-fidelity runs only to the survivors. BoTorch supports this through knowledge-gradient acquisitions that explicitly value information across fidelities.

The payoff is large because the cost gap between fidelities is often orders of magnitude — a simulation costs cents and a synthesis costs a day. A planner that burns high-fidelity budget on obviously-bad candidates is wasting the lab’s scarcest asset. The catch is that the correlation between fidelities must be real; if the cheap proxy is a poor predictor of the true measurement, multi-fidelity BO chases the proxy’s optimum and misleads you. Validate the fidelity correlation on a handful of paired runs before trusting it.

Constraints and mixed variable types

Physical labs have hard constraints: a formulation’s components must sum to one hundred percent, a temperature must stay under a decomposition threshold, a mixture must remain a single phase. Constrained BO handles these two ways. Known constraints prune the candidate set before optimization. Unknown constraints — where you only learn feasibility by running the experiment — get their own GP classifier, and the acquisition is weighted by the predicted probability of feasibility, so the planner avoids regions likely to fail. Mixed and categorical variables, discussed earlier, are handled by random-forest surrogates, specialized kernels, or tools like Gryffin and Atlas built for exactly this. The active-learning machinery generalizes cleanly from continuous knobs to “which of these forty candidate molecules do I test next,” which is the core of active learning for materials discovery.

A Worked Acquisition Decision and a Code Sketch

To make the mechanism concrete, here is an illustrative Expected Improvement calculation. The numbers are invented for clarity, clearly labelled, and not from any real dataset.

Suppose we are maximizing a reaction yield and the best measured yield so far is 0.70. The GP surrogate gives us three candidate experiments with these predictive means and standard deviations:

Candidate	Predicted mean	Predicted std	Character
A	0.72	0.02	high mean, low uncertainty
B	0.68	0.10	below best, high uncertainty
C	0.60	0.03	low mean, low uncertainty

For maximization, define the improvement over the incumbent best f-best as the standardized gap z equal to (mean minus f-best) divided by std. Expected Improvement equals std times the quantity z times the standard normal CDF of z plus the standard normal PDF of z.

Candidate A: gap is 0.72 minus 0.70 equals 0.02; z equals 0.02 divided by 0.02 equals 1.0. With CDF(1.0) about 0.841 and PDF(1.0) about 0.242, EI equals 0.02 times (1.0 times 0.841 plus 0.242), which is 0.02 times 1.083, about 0.0217.

Candidate B: gap is 0.68 minus 0.70 equals minus 0.02; z equals minus 0.02 divided by 0.10 equals minus 0.2. With CDF(minus 0.2) about 0.421 and PDF(minus 0.2) about 0.391, EI equals 0.10 times (minus 0.2 times 0.421 plus 0.391), which is 0.10 times 0.307, about 0.0307.

Candidate C: gap is 0.60 minus 0.70 equals minus 0.10; z equals minus 0.10 divided by 0.03 equals minus 3.33. Both CDF and PDF are tiny here, so EI is about 0.0004.

The ranking is B, then A, then C. This is the crucial lesson: candidate B has a lower predicted mean than the current best, yet EI ranks it first because its large uncertainty means a real chance of a big improvement. A greedy planner that only looked at the mean would pick A and never probe B. Expected Improvement explores on purpose. Change B’s std to 0.03 and its EI collapses below A’s — uncertainty, not mean, is doing the work.

Now watch what UCB does with the same three candidates, to see how the choice of acquisition changes the decision. With a moderately exploratory beta of 2.0, UCB scores mean plus 2.0 times std: A scores 0.72 plus 0.04 equals 0.76; B scores 0.68 plus 0.20 equals 0.88; C scores 0.60 plus 0.06 equals 0.66. UCB also ranks B first, but if we drop beta to 0.5 — a cautious, exploit-leaning setting — A scores 0.73 and B scores 0.73, a near tie, and any small change tips the decision to A. The same data, the same surrogate, and a different acquisition or a different beta produces a different experiment. That sensitivity is why the acquisition choice in Figure 4 is not a formality; it is the lever that sets how aggressively your campaign gambles reagents on uncertainty.

The batch case adds a second numeric lesson about diversity. Suppose the surrogate’s uncertainty is high across a whole ridge of candidate points that all look similar. Ranking by single-point EI would fill a 4-well plate with four points a few percent apart on that ridge, and after running them you would have four correlated measurements that collectively teach you about one region. A batch-aware objective like q-EI instead evaluates the joint improvement, so once the first ridge point is provisionally selected the marginal value of a second nearby point drops sharply, and the optimizer reaches for a point elsewhere in the space. The plate ends up sampling four distinct regions, and the surrogate improves everywhere rather than in one spot. Redundancy avoided is information gained.

A BoTorch-style code sketch

The pseudocode below follows the BoTorch and Ax API shape. It is illustrative and abbreviated, not a runnable script; real campaigns need standardization, input warping, and proper bounds handling.

# Illustrative BoTorch-style single-step proposal (not production code)
import torch
from botorch.models import SingleTaskGP
from botorch.fit import fit_gpytorch_mll
from gpytorch.mlls import ExactMarginalLogLikelihood
from botorch.acquisition import qLogExpectedImprovement
from botorch.optim import optimize_acqf

# train_X: measured experiment settings, normalized to the unit cube
# train_Y: measured objective, standardized (maximization)
model = SingleTaskGP(train_X, train_Y)          # GP surrogate, Matern 5/2 default
mll = ExactMarginalLogLikelihood(model.likelihood, model)
fit_gpytorch_mll(mll)                            # fit kernel hyperparameters

best_f = train_Y.max()
acqf = qLogExpectedImprovement(model=model, best_f=best_f)

# propose a batch of q=4 diverse experiments for the next plate
candidates, _ = optimize_acqf(
    acq_function=acqf,
    bounds=torch.stack([torch.zeros(train_X.shape[-1]),
                        torch.ones(train_X.shape[-1])]),
    q=4, num_restarts=10, raw_samples=256,
)
# hand `candidates` to the robotic platform; measure; append; refit

The structure mirrors the loop in Figure 1: fit the GP, build an acquisition (here the batch, numerically stable log-EI), optimize it to get q candidates, and hand those to the robot. For multi-objective you would swap in qLogExpectedHypervolumeImprovement; for cost-awareness you would wrap the acquisition in a cost model. The AI-planning layer that decides which objective to optimize in the first place is a separate concern, covered in our AI scientist hypothesis architecture.

One capability the sketch omits is warm-starting. A lab rarely optimizes a truly novel system from scratch; it has run related campaigns before. Transfer and meta-learning methods let a new campaign inherit priors from previous ones — a related material family, a similar reaction class — so the first few experiments are not blind. In practice this can be as simple as seeding the surrogate with historical data, or as sophisticated as a multi-task GP that shares structure across campaigns. The gain is front-loaded: warm-starting most improves the earliest, most uncertain phase of a campaign, which is exactly when a cold planner wastes the most reagents. The risk is negative transfer, where an ill-matched prior actively misleads the new campaign, so treat inherited priors as hypotheses the new data can overrule rather than as fixed truth.

Choosing an Acquisition Function

There is no universally best acquisition function; the right choice depends on objective count, noise, budget, and how aggressively you want to explore. Figure 4 gives a defensible default policy.

Figure 4: A practical decision flow from problem shape to acquisition-function choice.

Start with the number of objectives. Many objectives route straight to qEHVI over the Pareto frontier. For a single objective, branch on noise: heavy measurement noise favours Thompson sampling or a noise-robust EI variant, because deterministic EI can chase phantom improvements that are really just measurement scatter. For low-noise single-objective problems, branch on risk appetite: a safe exploit uses Expected Improvement, while pushing the frontier hard uses UCB with a large beta. When a single probe is exceptionally costly, entropy-based information search earns its extra compute by extracting maximum bits per experiment. Whatever the leaf, if the platform is parallel, wrap the choice in a batch variant — q-EI, qEHVI, or Thompson batching — to fill the plate.

Trade-offs, Gotchas, and What Goes Wrong

Bayesian optimization fails quietly, and a campaign can burn its whole budget before anyone notices the planner was misled. The most common failure is a bad prior. If the GP kernel or its length-scales are wrong — too smooth, too wiggly, badly scaled across dimensions — the surrogate’s uncertainty is miscalibrated and the acquisition function points at the wrong places. Always standardize inputs and outputs, and check that fitted length-scales are physically sane.

Non-stationarity is the second trap. A GP with a stationary kernel assumes the objective varies at the same rate everywhere. Real responses often have a flat plateau and a sharp cliff; a stationary kernel either over-smooths the cliff or over-fits the plateau. Input warping or non-stationary kernels help, but many teams never diagnose this and blame the method.

Noisy or irreproducible measurements break the incumbent-best logic. If your best-so-far is actually a lucky high-noise reading, EI and PI will keep hunting near a mirage. Use noise-aware acquisitions, replicate suspicious top results, and fit a noise term rather than assuming exact observations. Over-exploitation is the failure where the planner locks onto the first decent peak and stops exploring; UCB with too-small beta or PI with a tight margin both cause it, and you discover the global optimum was elsewhere only after the budget is gone.

Finally, dimensionality. Standard GP-BO degrades past ten to twenty dimensions because the volume to explore explodes and the cubic scaling in data bites. Options include trust-region methods like TuRBO, additive kernels that assume the objective decomposes, or dropping to random-forest and neural surrogates. And beware garbage-in: if the automated characterization mislabels a result, the planner faithfully optimizes toward the wrong target. The surrogate is only as good as the data feeding it, which is why measurement provenance and feature-store parity between offline and online data matter as much as the algorithm.

A subtler anti-pattern is treating the planner as infallible and switching off human judgment. BO optimizes exactly the objective you encode, so if the objective is a poor proxy for what you actually want — you maximize measured yield but ignore that half the high-yield conditions produce an unstable product — the planner will drive you confidently to a useless answer. Objectives should be reviewed by a domain expert, and a scientist should periodically inspect the surrogate’s picks for physical plausibility. Automation removes the tedium of choosing experiments; it does not remove the responsibility of choosing the right thing to optimize.

Practical Recommendations

Start simple and only add complexity when the data demands it. For a first single-objective campaign, use a Gaussian process with a Matern 5/2 kernel and Expected Improvement — it is well understood, has a closed form, and rarely embarrasses you. Standardize every input to the unit cube and every output to zero mean and unit variance before fitting; scale bugs cause more failed campaigns than algorithm choice does. Seed the campaign with a small space-filling design, roughly five to ten points per dimension via Sobol or Latin hypercube, so the first surrogate is not fitting noise.

Match the batch size to your physical throughput. If the robot runs 24 wells, propose 24 diverse points with q-EI or Thompson batching rather than one point at a time. Log the fitted length-scales and predictive uncertainty every round and actually look at them — a length-scale pinned at its bound or an uncertainty that never shrinks is a red flag. Replicate your best result before you trust it. Budget in resource units, not experiment counts, and make the planner cost-aware if your experiments differ in expense.

Use the mature libraries: BoTorch and Ax for general and multi-objective work, Gryffin or Atlas for categorical chemistry spaces, and Olympus benchmark surfaces to sanity-check a planner before you point it at real reagents.

[ ] Standardize inputs and outputs before every fit.
[ ] Seed with a Sobol or Latin-hypercube design.
[ ] Default to GP plus Expected Improvement for one objective; qEHVI for many.
[ ] Set batch size q to the platform’s parallelism.
[ ] Make the acquisition cost-aware when experiment costs differ.
[ ] Log length-scales and posterior uncertainty each round.
[ ] Replicate the incumbent best before declaring victory.
[ ] Switch to Thompson or noisy EI under heavy measurement noise.

Frequently Asked Questions

What is Bayesian optimization in autonomous experiments?

It is a sequential, model-based method for choosing which expensive physical experiment to run next. A probabilistic surrogate, usually a Gaussian process, predicts the objective and its uncertainty across the design space from the data measured so far. An acquisition function scores each candidate, and the planner runs the highest-scoring one. Inside a self-driving lab it is the decision layer that turns measured results into the next experiment, minimizing how many costly runs you need to find a strong candidate.

How does an acquisition function balance exploration and exploitation?

It combines the surrogate’s predicted value with its predicted uncertainty into one score. Exploitation favours points with high predicted value; exploration favours points with high uncertainty. Expected Improvement blends both implicitly by scoring the expected gain over the current best. Upper Confidence Bound makes the balance explicit through a beta multiplier on the uncertainty. A point can win either by looking good on average or by being uncertain enough that it might be excellent, which is how the planner avoids getting stuck on the first decent result.

When should I use batch parallel Bayesian optimization?

Whenever the physical platform can run several experiments simultaneously — a 24- or 96-well plate, or parallel reactors. Sequential BO proposes one point per round and wastes idle capacity. Batch methods such as q-EI, local penalization, and Thompson batching propose q points that are individually promising yet mutually diverse, so the plate is not filled with near-identical runs. Matching the batch size to the hardware’s parallelism is usually the single biggest wall-clock speedup available to a self-driving lab.

Is a Gaussian process always the best surrogate model?

No. Gaussian processes give excellent calibrated uncertainty for small datasets and up to roughly ten to twenty continuous dimensions, but their cost scales cubically with data and they handle categorical variables poorly. For high dimensions, large datasets, or discrete choices like which solvent or ligand, random-forest surrogates, Bayesian neural networks, deep ensembles, or chemistry-specific tools like Gryffin often work better. The surrogate must supply calibrated uncertainty; the exact model is a fit-for-purpose choice, not a fixed rule.

How does cost-aware Bayesian optimization work?

It accounts for the fact that experiments differ in expense. Instead of maximizing raw acquisition value, a cost-aware planner maximizes acquisition value per unit cost — Expected Improvement divided by estimated cost is the canonical form. This makes the planner prefer cheap, informative experiments and only commit to an expensive one when the expected payoff justifies the price. When cost is itself unknown, a second surrogate can model it, and the planner optimizes expected improvement per expected cost, budgeting in resource units rather than experiment counts.

What are the most common ways Bayesian optimization fails?

Bad priors with mis-scaled or wrongly-shaped kernels give miscalibrated uncertainty and point the planner at the wrong regions. Non-stationarity breaks a stationary kernel when the objective has both flat plateaus and sharp cliffs. Noisy or irreproducible measurements corrupt the incumbent-best logic and send the planner chasing a mirage. Over-exploitation locks onto the first peak and skips the global optimum. High dimensionality degrades standard GP-BO. Most of these fail quietly, so log uncertainties and replicate top results.

Bayesian Optimization for Autonomous Experiments: The Planner Inside a Self-Driving Lab (2026)