The framing most teams bring to active learning is wrong before the first experiment runs. They think of it as a way to build a bigger training set faster. It is not. Active learning is a strategy for deciding what information is worth collecting given a fixed budget and a specific objective. That distinction changes which acquisition function you reach for, and it changes what "good" looks like at the end of a campaign.
At Matforgelab we have run active learning loops on synthesis planning problems across battery electrolytes, structural alloys, and thermoset formulations. The failure mode is almost never a bad surrogate model. It is a mismatch between the acquisition function and what the team actually wants to know.
What the Surrogate Does and Does Not Do
A Gaussian process surrogate built on 40 synthesis-characterization pairs will give you a mean prediction and a posterior variance estimate for any candidate composition. The mean tells you where the model thinks a property lands. The variance tells you how uncertain the model is, which in a Gaussian process correlates with how sparsely sampled the neighbourhood is.
Neither of those alone is sufficient to plan the next experiment. You need to combine them according to your goal. If your goal is to maximize yield-under-fatigue-load for an Al-Cu-Mg ternary, you care about finding the global maximum. If your goal is to verify that a candidate is above a threshold before moving to scale-up, you care about classification accuracy near the threshold. If your goal is to build a general-purpose property model across the whole composition space, you care about reducing total uncertainty everywhere.
These are three different goals. They require three different acquisition functions. Confusing them is the single most common reason an active learning campaign underperforms brute-force random sampling.
Three Acquisition Strategies and When Each Applies
Expected Improvement for Global Optimization
Expected Improvement (EI) integrates over the probability distribution of improvement over the current best observed value. It naturally balances exploitation (sampling near known good regions) against exploration (sampling uncertain regions where you might find something better). For a synthesis campaign trying to maximize ionic conductivity in a solid electrolyte, EI is usually the right default.
The failure case for EI is a composition space with many local optima at the scale of your grid resolution. EI will tunnel into the first strong local optimum it finds and refuse to leave. When we first ran a Li-La-Zr-O oxide campaign early on, we saw exactly this: 28 of 35 active learning queries landed in a tight cluster around an LLZO composition we already understood well. The model was confident and the expected improvement calculations kept pointing back there. We added a Thompson sampling component to force diversification, which helped, but it also slowed convergence toward the actual maximum. There is no free lunch.
Probability of Improvement for Threshold Classification
Probability of Improvement (PI) asks: what is the chance the next sample at this candidate exceeds the current best by at least epsilon? When epsilon is small, PI behaves greedily. When epsilon is set to a meaningful threshold, say, a minimum conductivity spec for a battery separator, PI concentrates queries near the decision boundary where classification uncertainty is highest.
This is the right choice when a binary pass/fail decision drives the value of the campaign. A team qualifying a polymer binder for a cathode slurry does not need to know whether the best formulation achieves 95 MPa tensile or 97 MPa. They need to know which formulations clear 80 MPa. PI directs experiments toward resolving that boundary efficiently.
Maximum Variance for Model Building
Pure variance-based sampling, sometimes called IVAR or maximum entropy sampling in the GP literature, ignores the objective function entirely and queries wherever the model is most uncertain. This is appropriate when the deliverable is a calibrated, broadly-applicable property model rather than a single optimal composition. If a team is building an internal predictive tool that will serve dozens of future formulation decisions, you want the model to be accurate everywhere, not just near one peak.
The cost: maximum variance sampling is slow to find optima. If you have 30 experiments to spend, pure exploration will rarely give you a high-performing candidate by the end of the campaign. We do not recommend it unless the primary output is the model itself.
Batch Acquisition: The Reality of Lab Throughput
Real synthesis campaigns run in batches. You do not synthesise one composition, measure it, then query the acquisition function again. You plan a batch of 8 or 12 compositions, hand them to the lab, wait a week, and then update. This breaks the sequential assumption that most active learning derivations rest on.
There are two common approaches for batch acquisition. The first is a greedy approximation: select the top-k candidates by your acquisition function, but after each selection, update the GP as if the selected candidate had been observed with the predicted mean as its outcome. This is sometimes called the "kriging believer" approach. It works reasonably well in practice and it is cheap to compute.
The second approach uses a determinantal point process or similar diversity penalty to spread batch selections across the composition space while still weighting toward high-acquisition regions. This tends to outperform greedy selection in high-dimensional spaces where the acquisition landscape has broad flat peaks.
In practice the difference between these two approaches is smaller than the choice of acquisition function. We have found that greedy kriging believer batches of 8 to 12, with EI as the acquisition function, converge well on optimization tasks within 4 to 5 batch rounds when the composition space is below 5 or 6 dimensions. Beyond 8 dimensions, the problem geometry changes enough that you need to think carefully about dimensionality reduction before the GP fitting stage.
The Kernel Choice Is Not a Detail
Gaussian process surrogates for composition-property relationships are only as good as the kernel function encoding your assumptions about how properties vary with composition. The default RBF (squared exponential) kernel assumes properties vary smoothly and infinitely differentiably. This is a strong assumption for materials.
For alloy composition spaces, a Matern 5/2 kernel is almost always more appropriate. It assumes properties are twice differentiable but not more, which better matches the behaviour of yield strength or hardness across compositional gradients near phase boundaries. For polymer formulations with strong nonlinear interactions between components, we often add an additive component to the kernel to capture linear compositional effects separately from interaction effects.
We are not saying that the choice of kernel is the most important decision in the workflow. The acquisition function and the experimental design of the initial dataset matter more. But poor kernel choice will cap your convergence rate regardless of how well you run the rest of the campaign, and it is one of the easier things to tune given access to cross-validation data.
Initializing the Campaign: The First 15 Are Expensive
Active learning cannot start from zero data. You need an initial dataset to fit the first surrogate. The quality of that initial dataset sets the ceiling on how fast the campaign converges in the first several rounds.
Two strategies dominate in materials synthesis planning. Latin hypercube sampling (LHS) spreads initial points evenly across the composition space in a space-filling design. Simplex lattice design (for simplex composition constraints) does the same with the correct constraint geometry. Both are substantially better than grid sampling or convenient round-number compositions, which is what most teams default to when they start an empirical campaign without a systematic initial design.
A practical rule: 2 to 3 times the number of composition dimensions is a reasonable minimum initial dataset size. For a 5-component alloy system, that means 10 to 15 carefully placed initial experiments before you let the active learner run. This feels expensive upfront. It pays off by round 3 when your GP variance is low enough to make EI recommendations meaningful rather than arbitrary.
Where Active Learning Adds the Most Value
We see the largest impact in two scenarios. The first is when synthesis is the bottleneck: experiments cost $800 to $3,000 each, cycle time is 3 to 5 days, and the team has 40 to 60 experiments to spend before a go/no-go decision. Active learning typically finds the optimum in 50 to 70 percent of the budget that random sequential sampling would require. That translates to months, not percentages.
The second scenario is when the property landscape is rough and non-intuitive, meaning expert heuristics do not reliably point toward the optimum. For a new class of polymer electrolyte where the mechanism linking monomer composition to ionic conductivity is not well understood, expert intuition is a weak prior. The active learner starts from the same weak prior but is systematic about where to resolve uncertainty next.
Where active learning adds less value: tasks where 80 percent of the space is ruled out by physics-based constraints before experimentation begins, and tasks where the objective is so noisy that GP variance estimates are dominated by measurement noise rather than composition-space uncertainty. In those cases, careful experimental design of a fixed-size batch often outperforms an iterative active learning loop.
The technology works best when the decision about what to measure next is genuinely hard, when the budget is genuinely limited, and when you have a clear enough objective to translate it into an acquisition function. That combination is exactly the situation most materials R&D campaigns find themselves in.