20 Cycles, 4 Weeks: A Materials Optimization Case Study

Thermal conductivity was the optimization target for a silicone composite with two fillers: boron nitride platelets and aluminum oxide particles. In early 2026, a four-person materials group joined Discovery Loop through our early-access program. They varied total filler loading fraction, the BN-to-AlO ratio, and silane treatment concentration applied to BN platelets before blending. The team ran twenty experiments over four weeks. This is the resulting cycle log.

The Initial Design

The team supplied six earlier manual experiments: three total loading fractions at 40, 50, and 60 percent by volume, two BN-to-AlO ratios, and no surface treatment. Thermal conductivity ranged from 1.8 to 3.1 W/mK. Their plan was to extend the loading-fraction grid upward, assuming conductivity would rise monotonically with filler content.

After we loaded those results into the hypothesis tree, the model assigned more predictive weight to the BN-to-AlO ratio than to loading fraction alone across the tested range. Published work on filler synergy supported that estimate, though the team's data did not yet show the relationship clearly. Its highest-ranked first run used a high BN ratio at 55 percent loading, a condition they had not tested. They approved it as cycle 1.

Cycles 1 Through 5: Mapping the Space

Cycle 1 produced 3.6 W/mK, then the dataset's top result. The model shifted toward high BN ratios and selected cycles 2 and 3 as nearby tests: slightly higher loading at 58 percent, followed by a higher BN ratio at the same 55 percent loading. Results fell between 3.4 and 3.7 W/mK, supporting the update.

Cycles 4 and 5 tested surface treatment, a variable absent from the prior data. Its covariance estimate therefore came entirely from the assumed length-scale prior, making both runs genuinely exploratory. The team approved that uncertainty-reduction spend rather than expecting an immediate gain. Cycle 4 reached 3.8 W/mK at low treatment concentration; cycle 5 reached 4.2 W/mK at medium concentration. The posterior then moved sharply toward treated, high-BN conditions, extending leaf nodes in that region.

Cycles 6 Through 12: Exploitation and One Override

At cycle 6, most probability mass occupied a subspace with BN ratio above 0.70, total loading between 55 and 65 percent, and medium to high surface treatment. Cycles 6 through 12 varied those conditions systematically. As the posterior narrowed, each proposed change became tighter.

The team overrode one recommendation in this interval. For cycle 9, they fixed loading at 35 percent and treatment at a high level to test whether treatment affected conductivity independently of loading. The model had ranked this condition low because of the loading trend. It yielded 2.9 W/mK, below the prior manual baseline of 3.1 W/mK.

This is a useful role for the loop: a researcher can run a hypothesis the model has deprioritized to answer a structural physics question. The result showed that surface treatment does not offset low filler loading on its own. We incorporated that observation, raised the lower loading bound, and returned to high-loading, high-treatment conditions for cycles 10 through 12.

Cycles 13 Through 20: Refinement

Cycles 13 through 20 stayed in a narrow neighborhood: BN ratio changed in 0.05 increments from 0.72 to 0.82, total loading ranged from 58 to 63 percent, and treatment concentration advanced in 5 percent increments. With the broad region established, these cycles searched for its local optimum.

Cycle 17 reached 4.9 W/mK, the maximum in the full 26-experiment dataset, including the six earlier manual runs. Cycles 18, 19, and 20 tested adjacent settings and returned 4.6, 4.8, and 4.7 W/mK, respectively, all within cycle 17's measurement uncertainty. Outside the loop, the team ran two confirmation samples at the cycle 17 condition. Both measured between 4.7 and 5.1 W/mK.

Lessons From the Log

The sequential log exposes three clear patterns.

The model reached the useful region sooner than the planned manual grid. That sequential approach would have swept loading first and examined surface treatment only afterward, likely reaching treated, high-BN conditions around cycle 14 or 15 instead of cycle 5.

Cycles 4 and 5 produced the project's highest information yield as a pair. With no prior treatment data, the model spent two runs reducing real uncertainty instead of pursuing small performance gains. That exploration immediately improved the search.

The cycle 9 override also paid off. The loop does not remove researcher judgment. It takes on routine low-load decisions while preserving an override for a question the model cannot rank effectively, making that human intervention more targeted.

What We Would Change

Repeating the project, we would seed the prior data differently. The initial six experiments covered loading fraction and BN ratio but omitted surface treatment entirely. As a result, cycles 4 and 5 had to establish that variable's relevance instead of starting with one or two characterized treatment levels.

For future loop initialization, cover every categorical or binary variable at least once before adding density to continuous-variable grids. Resolving qualitative dimensions early leaves the model free to spend quantitative exploration where it affects the outcome.