Agentic Resource Allocation for Batch Multi-Objective Bayesian Optimization in Autonomous Materials Discovery
Authors: Robert Robinson, Shakti Prasad Padhy, Sushant Sinha, Sk Md Ahnaf Akif Alvi, Juan Florez Coronel, Brent Vela, Trevor Hastings, Douglas Allaire, Raymundo Arróyave
Organizations: Department of Materials Science and Engineering, Texas A&M University, College Station, 77843, TX, USA · J. Mike Walker ’66 Department of Mechanical Engineering, Texas A&M University, College Station, 77843, TX, USA · Wm Michael Barnes ’64 Department of Industrial and Systems Engineering, College Station, 77843, TX, USA
The discovery and development of advanced materials is a challenging process constrained by the high time and monetary costs of synthesis, processing, and characterization. The underlying design spaces can be enormous, often with multiple competing objectives. Bayesian optimization (BO) provides a principled approach for efficiently navigating such spaces, but most workflows rely on fixed exploration-exploitation policies that lack the capacity to adapt to shifting constraints in dynamic campaigns typical of self-driving laboratories. In this work, we develop a multi-objective BO framework for alloy design under resource constraints, benchmarking strategies for adaptive policy tuning at each iteration. Our evaluation covers a septenary refractory high-entropy alloy (RHEA) system focused on maximizing melting temperature and minimizing density, and an Fe-Co-Ni-based soft magnetic alloy system targeting saturation magnetization, coercivity, and hardness. We compare an exploitation-focused strategy, a fixed mixed exploratory/exploitative policy, and two distinct LLM-based adaptive strategies with different approaches to batch allocation and campaign signal interpretation, evaluated across baseline and mid-campaign resource event conditions including budget reductions, timeline cuts, and combined disruptions. Our results show that mixed allocation strategies accumulate substantially more mutual information than the exploitation-focused baseline at a proportionally smaller cost to hypervolume and optimization speed, with adaptive strategies outperforming a fixed-mixed allocation policy by adjusting their allocation in response to both evolving campaign statistics and resource constraints. These findings suggest that adaptive resource allocation offers a favorable tradeoff for materials discovery campaigns in reducing predictive uncertainty on Pareto-optimal compositions.
Figures & tables
Figure 1: Overview of the developed LLM-guided resource allocation framework for batch multi-objective Bayesian optimization. The optimization loop is shared across all methods: surrogate fitting, state estimation, allocation-option generation, batch evaluation, and update. The proposed agent selects among exploit-explore batch allocations based on the current decision state, whereas the baselines use fixed qEHVI-only or qUCB-only allocations.
System
ntrees
max_depth
min_samples_split
min_samples_leaf
RHEA
53
56
3
1
Fe-Co-Ni
134
113
7
1
Table 1: Optimal RFR hyperparameters for each alloy system selected by ARMOTE-MO.
Figure 2: Parity plots for the ground truth random forest regression models. Panels (a)–(c) show the Fe–Co–Ni–V–Mo–C–Nb–Ti–Si soft magnetic system with predicted versus measured values for Ms (T, R2=0.97 ), logHc ( log(A/m) , R2=0.91 ), and logHV ( log(HV) , R2=0.77 ). Panels (d)–(e) show the Ti–V–Nb–Mo–Hf–Ta–W refractory system with predicted versus measured values for Tm (K, R2=0.89 ) and ρ (g/cm 3 , R2=0.93 ). The dashed line indicates perfect agreement.
Experiment
System
Seeds
Iters
q
ninit
Event(s)
MP-D-MAIN
RHEA
25
55
5
5
None
MP-D-BUDGET
RHEA
25
55
5
5
Budget cut at iter. 15
MP-D-TIME
RHEA
25
55
5
5
Time cut at iter. 15
MP-D-DUAL
RHEA
25
55
5
5
Time cut at iter. 5; budget cut at iter. 7
Magnets-MAIN
Fe-Co-Ni
25
50
5
15
None
Magnets-BUDGET
Fe-Co-Ni
25
50
5
15
Budget cut at iter. 18
Table 2: Summary of all experiment configurations. Iters denotes the maximum number of BO iterations; campaigns with resource events terminate earlier when a constraint is exhausted. q denotes batch size. ninit denotes the number of initial compositions shared across all strategies within each seed.
Figure 3: Ground-truth design and objective spaces for both systems. (a) RHEA system: training data and Pareto-optimal compositions shown in the septenary composition simplex. (b) RHEA objective space: density vs. melting temperature for all compositions, with Pareto-optimal points highlighted. (c) Fe-Co-Ni magnet system: training data and Pareto-optimal compositions projected onto the Fe–Co–Ni ternary, with Pareto-optimal compositions highlighted. (d–f) Fe-Co-Ni magnet objective space: pairwise projections of the three objectives, showing the multi-clustered structure of the Pareto front.
Figure 4: Baseline campaign results for both alloy systems with no resource events ( n=25 seeds per strategy). a–b: Hypervolume convergence; lines show per-strategy means and shaded bands indicate 95% confidence intervals. c–d: Per-iteration mutual information across seeds; shaded bands are 95% confidence intervals. Left column: RHEA system ( q=5 , 55 iterations). Right column: Fe-Co-Ni magnet system ( q=5 , 50 iterations). Strategies: Adaptive- β (blue), Rule-Swap (purple), qEHVI (orange), Fixed Mixed (green).
Figure 5: The information–optimization tradeoff across all eight experimental configurations ( n=25 seeds per strategy). (a–b) Cumulative mutual information over the baseline campaign for the RHEA and Fe-Co-Ni magnet systems; lines are per-strategy means and shaded bands are 95% confidence intervals. (c) Cumulative MI gain plotted against AUC-HV cost, both expressed as a percentage change relative to the qEHVI baseline of the same experiment, which sits at the origin. Marker colour denotes strategy, and each experiment contributes one marker per mixed-allocation strategy; filled markers are the RHEA system and translucent markers the Fe-Co-Ni magnet system. The dotted line marks an equal trade, where the percentage of information gained equals the percentage of convergence speed given up. Every mixed-allocation point lies far above that line. End-of-campaign cumulative MI means are annotated at the right edge of (a) and (b); those values for all eight configurations, with their standard errors and alongside final hypervolume and AUC-HV, are tabulated in C.1 . Pairwise strategy comparisons reported anywhere in this work use the two-sided Wilcoxon signed-rank test on per-seed differences matched by shared random seed, with no multiple-comparison correction applied across the six strategy pairs within an experiment; the full matrix of p -values is given in C.2 .
Figure 6: Hypervolume convergence for all six event experiments. Mean ± 95% confidence interval across n=25 seeds. Dashed vertical lines mark the iteration(s) at which resource events were applied. Top row : RHEA system ( q=5 ). Bottom row : Fe-Co-Ni magnet system ( q=5 ). From left to right: budget-event, time-event, and dual-event configurations. Strategies: Adaptive- β (blue), Rule-Swap (purple), qEHVI (orange), Fixed Mixed (green).
System
Experiment
Strategy
Batches to 50% HV
Batches to 80% HV
Mean ± 95% CI
Mean ± 95% CI
% reached
Fe-Co-Ni
Baseline
Adaptive- β
2.7±0.4
29.7±4.6
96%
Rule-Swap
2.5±0.4
27.4±4.1
92%
qEHVI
2.4±0.3
26.3±4.2
96%
Fixed Mixed
2.9±0.6
28.3±3.9
96%
Budget event
Adaptive- β
2.6±0.4
20.1±2.5
64%
Table 3: Batches required to reach 50% and the upper milestone hypervolume threshold (mean ± 95% CI; n=25 seeds per strategy). All seeds reach the 50% milestone in every experiment. Upper milestone thresholds differ by system, 80% for the Fe-Co-Ni magnet problem; 90% for the RHEA problem. The % column reports the fraction of seeds reaching the upper milestone.
Figure 7: Adaptive- β (blue, circles) and Rule-Swap (purple, diamonds) β selection across all eight experimental configurations ( n=25 seeds per strategy). Top row (a–d) : RHEA system. Bottom row (e–h) : Fe-Co-Ni magnet system. Columns from left to right show the baseline, budget cut, time cut, and time-plus-budget cut configurations. Lines show per-seed means; shaded bands are 95% confidence intervals. Dashed vertical lines mark the iterations at which resource events were applied. The dotted horizontal line marks β=2.0 , the minimum selectable value.
Figure 8: Exploration candidate allocation ( nexp ) for Adaptive- β and Rule-Swap across all six event experiments. Mean across n=25 seeds; shaded bands are 95% confidence intervals. Dashed vertical lines mark the iteration at which each resource event was applied; colors distinguish event type (pink: first event; orange: second event in dual-event experiments). The dotted horizontal line marks the balanced allocation. Top row : RHEA system. Bottom row : Fe-Co-Ni magnet system. From left to right: budget-event, time-event, and dual-event configurations.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Hyperparameter
Search range
Scale
ntrees
[10, 150]
integer, uniform
max_depth
[10, 150]
integer, uniform
min_samples_split
[2, 20]
integer, uniform
min_samples_leaf
[1, 20]
integer, uniform
Appendix
Table 4: Hyperparameter search spaces explored by ARMOTE-MO for both alloy systems. Ranges are inclusive. All other random forest settings, including max_features , were left at their scikit-learn defaults. The selected configurations are reported in Table 1 .
System
Experiment
Strategy
Cumulative MI (nats)
Final HV
AUC-HV
RHEA
Baseline
Adaptive- β
19.54±0.41
0.9967±0.00019
51.59±0.092
Rule-Swap
17.20±0.50
0.9970±0.00018
51.79±0.097
Fixed Mixed
15.48±0.34
0.9971±0.00017
50.98±0.093
qEHVI
14.23±0.45
0.9983±0.00018
52.02±0.077
RHEA
Budget cut
Adaptive- β
12.29±0.38
0.9843±0.00071
21.87±0.078
Rule-Swap
10.93±0.43
0.9861±0.00048
21.99±0.086
Appendix
Table 5: Cumulative mutual information, final normalized hypervolume, and area under the hypervolume curve for every system, experiment, and strategy (mean ± SEM, n=25 seeds).
System
Experiment
Strategy pair
p (final HV)
p (AUC-HV)
p (cum. MI)
RHEA
Baseline
Adaptive- β vs. Rule-Swap
0.090
0.006 **
<0.001 ***
Adaptive- β vs. Fixed Mixed
0.080
<0.001 ***
<0.001 ***
Adaptive- β vs. qEHVI
<0.001 ***
<0.001 ***
<0.001 ***
Rule-Swap vs. Fixed Mixed
0.958
<0.001 ***
0.006 **
Rule-Swap vs. qEHVI
<0.001 ***
<0.001 ***
<0.001 ***
qEHVI vs. Fixed Mixed
<0.001 ***
<0.001 ***
0.026 *
Appendix
Table 6: Pairwise strategy comparisons for final normalized hypervolume, AUC-HV, and cumulative mutual information ( n=25 seeds per strategy). Two-sided paired Wilcoxon signed-rank, uncorrected; markers denote * p<0.05 , ** p<0.01 , *** p<0.001 .
Advanced materials discovery increasingly relies on machine learning and Bayesian optimization to explore large discrete design spaces under limited evaluation budgets. However, conventional Bayesian optimization (BO) can become inefficient as candidate spaces grow, often evaluating low-value regions before reaching informative areas. We propose an active-learning (AL)-guided adaptive search-space refinement framework combined with multi-objective BO to accelerate materials optimization while preserving Pareto-relevant regions. We evaluate the approach on CH4/N2 separation in covalent-organic frameworks and pressure-vessel design with material-direction stress components and thickness objectives. Results show that the AL-guided refinement reduces the candidate space by approximately half while preserving more than 99 percent of the original hypervolume. The reduced-space strategy improves early convergence and cumulative Pareto-front discovery from the BO, demonstrating efficient large-scale materials optimization across constrained autonomous materials discovery settings.
Alexandros Ntagiantas, Panagiotis Tsilimidos, George Giannakopoulos +2
National Centre for Scientific Research “Demokritos”
Autonomous laboratories automate experimental execution, but a campaign must also decide which recovery pathway merits optimization. We formulate this as a sequential decision problem with a discrete pathway-identification stage and a continuous within-pathway optimization stage under heterogeneous experimental costs. Our implementation, Coactive learning, combines a cost-sensitive Bayesian hypothesis-discrimination policy motivated by EC2 (Golovin et al., 2010) with Gaussian-process Bayesian optimization (Srinivas et al., 2010). Under explicitly stated assumptions, the expected spend of one fixed-budget campaign attempt is bounded by the expected pathway-identification cost plus the capped within-pathway optimization budget. We evaluate the method on synthetic benchmarks constrained by selected results reported for PNNL's CICERO selective-precipitation study (Ritchhart et al., 2026). The method performs comparably to an oracle-pathway Bayesian-optimization reference and to a strong split-plate baseline that discriminates pathways with its first plate, without receiving an oracle label for the correct pathway. It is given a candidate hypothesis space and a diagnostic likelihood model. On an NdFeB-inspired instance, it avoids the simulated penalty of a commit-first baseline that initially selects a plausible but inferior hydroxide pathway. This hypothetical wrong-first-commitment scenario is motivated by the hydroxide-oxalate performance contrast reported by CICERO. We characterize the sensitivity of these conclusions to the assumed cost model. The code and benchmark are open source.
Materials and chemical optimization must choose the next synthesis, characterization, or high-fidelity simulation from sequential observations under a limited budget. Large language models can use task knowledge to propose adjustments, but composing surrogates, acquisition functions, and search strategies is constrained by compatibility, parameter dependencies, and evidence requirements. We present Agentic-MatriBO, a capability-compiled agentic Bayesian optimization (BO) framework that represents BO components and selection conditions as composable capabilities. A language policy proposes BO recipes and candidate plans; a compiler checks proposals against the registered capability catalog; and a deterministic execution layer validates and commits candidates. Oracle observations are stored in research memory and inform later rounds. We compare Agentic-MatriBO with eight baselines on four offline finite candidate pools---COF, PCE10, Buchwald--Hartwig (BH), and Suzuki--Miyaura---using five seeds per condition; results are reported descriptively. Across the four tasks, Agentic-MatriBO has the best mean endpoint quality after orienting each objective so that higher is better. On COF, it reaches the tabulated optimum in all five seeds both in the end-to-end comparison and after shared initialization, where it also retains an advantage in cumulative quality. Its normalized area under the best-so-far quality curve during the BO stage (nAUC-BO) is substantially higher on COF, near parity with the best baseline on PCE10 and BH, and lower on Suzuki. COF ablations show that removing the capability compiler or evidence gate coincides with lower cumulative quality and fewer exact-optimum hits. The framework provides an inspectable interface for evidence-conditioned BO adaptation.