Multi-objective Bayesian Optimization with Human-in-the-Loop for Flexible Neuromorphic Electronics Fabrication
Pith reviewed 2026-05-18 09:12 UTC · model grok-4.3
The pith
Incorporating failed experiments through human feedback accelerates multi-objective Bayesian optimization for fabricating flexible neuromorphic electronics.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The central discovery is that a human-in-the-loop framework for incorporating failed experiments into the MOBO workflow accelerates optimization by reducing the number of experimental rounds required to identify Pareto-optimal photonic curing conditions for flexible metal-insulator-metal capacitors with tailored aluminum oxide dielectrics.
What carries the argument
The human-in-the-loop framework that supplies failed experimental outcomes as additional data to the multi-objective Bayesian optimization model so the search avoids unproductive parameter regions.
If this is right
- Fewer experimental iterations are needed to reach usable processing conditions when failures are modeled explicitly.
- The framework applies to any multi-objective experimental problem that combines many inputs with high failure rates.
- Examination of the resulting Pareto set shows how to adjust dielectric properties for neuromorphic requirements.
- Shapley analysis ranks the relative influence of each curing input on the observed electrical outcomes.
Where Pith is reading between the lines
- Treating failures as data could lower material and time waste in other lab-scale flexible electronics processes.
- The same loop might speed development of solution-processed materials beyond aluminum oxide.
- If additional performance metrics such as mechanical bendability are added, the optimization could better match full device needs.
- Scaling the approach to automated closed-loop fabrication would require reliable ways to detect and log failures in real time.
Load-bearing premise
That failed experiments can be added to the model in a way that improves the search instead of biasing it toward already-tried regions.
What would settle it
A head-to-head comparison of the same optimization task run once with the human-in-the-loop failure incorporation and once without it, measuring the number of rounds until comparable Pareto fronts are reached.
read the original abstract
Neuromorphic computing hardware enables edge computing and can be implemented in flexible electronics for novel applications. Metal oxide materials are promising candidates for fabricating flexible neuromorphic electronics, but suffer from processing constraints due to the incompatibilities between oxides and polymer substrates. In this work, we use photonic curing to fabricate flexible metal-insulator-metal capacitors with solution-processible aluminum oxide dielectric tailored for neuromorphic applications. Because photonic curing outcomes depend on many input parameters, identifying an optimal processing condition through a traditional grid-search approach is unfeasible. Here, we apply multi-objective Bayesian optimization (MOBO) to determine photonic curing conditions that optimize the trade-off between desired electrical properties of large capacitance-frequency dispersion and low leakage current. Furthermore, we develop a human-in-the-loop (HITL) framework for incorporating failed experiments into the MOBO machine learning workflow, demonstrating that this framework accelerates optimization by reducing the number of experimental rounds required. Once optimization is concluded, we analyze different Pareto-optimal conditions to tune the dielectrics properties and provide insight into the importance of different inputs through Shapley Additive exPlanations analysis. The demonstrated framework of combining MOBO with HITL feedback can be adapted to a wide range of multi-objective experimental problems that have interconnected inputs and high experimental failure rates to generate usable results for machine learning models.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript applies multi-objective Bayesian optimization (MOBO) to photonic curing parameters for solution-processed aluminum oxide dielectrics in flexible metal-insulator-metal capacitors targeting neuromorphic use. The two objectives are large capacitance-frequency dispersion and low leakage current. A human-in-the-loop (HITL) component is introduced to feed failed experiments back into the MOBO surrogate and acquisition function, with the central claim that this reduces the number of experimental rounds needed to reach usable Pareto fronts. After optimization, selected Pareto points are characterized and SHAP analysis is performed to rank input-parameter importance.
Significance. If the HITL mechanism can be shown to accelerate search without biasing the surrogate toward already-sampled regions, the approach would be useful for any high-failure-rate multi-objective fabrication task. The combination of MOBO with post-hoc SHAP interpretability is a modest but practical contribution for experimentalists who must balance electrical metrics under processing constraints.
major comments (2)
- [Results / Experimental workflow] The abstract and results sections assert that the HITL framework 'accelerates optimization by reducing the number of experimental rounds required,' yet no baseline MOBO run (without HITL), no count of rounds saved, no error bars on convergence curves, and no statistical comparison are reported. Without these data the central claim cannot be evaluated.
- [Methods / HITL framework] The precise rule for incorporating failed experiments (constraint, penalized objective, separate failure classifier, or simple discard) is not stated in the methods or algorithm description. Consequently it is impossible to determine whether any observed speedup arises from genuine avoidance of high-failure regions or from implicit narrowing of the search space to previously tried parameter values.
minor comments (2)
- [Figures] Figure captions and axis labels should explicitly state the number of independent fabrication runs and the failure rate observed in each MOBO iteration.
- [Introduction / Objectives] The two chosen objectives (capacitance-frequency dispersion and leakage current) are reasonable but the manuscript should briefly justify why additional neuromorphic-relevant metrics (e.g., endurance or retention) were not included.
Simulated Author's Rebuttal
We thank the referee for their constructive and detailed review. We address each major comment below, indicating where revisions will be made to strengthen the manuscript.
read point-by-point responses
-
Referee: [Results / Experimental workflow] The abstract and results sections assert that the HITL framework 'accelerates optimization by reducing the number of experimental rounds required,' yet no baseline MOBO run (without HITL), no count of rounds saved, no error bars on convergence curves, and no statistical comparison are reported. Without these data the central claim cannot be evaluated.
Authors: We agree that a direct baseline comparison against standard MOBO without HITL would allow a more rigorous quantification of the acceleration. Because the photonic curing process exhibits a high failure rate, a pure MOBO run without failure feedback was not performed experimentally. In the revised manuscript we will add a retrospective analysis that uses the trained surrogate models to simulate a non-HITL baseline, reporting estimated rounds to reach comparable Pareto quality together with variability estimates derived from multiple acquisition-function seeds. We will also tabulate the number of failed experiments encountered and avoided during the actual HITL campaign. revision: partial
-
Referee: [Methods / HITL framework] The precise rule for incorporating failed experiments (constraint, penalized objective, separate failure classifier, or simple discard) is not stated in the methods or algorithm description. Consequently it is impossible to determine whether any observed speedup arises from genuine avoidance of high-failure regions or from implicit narrowing of the search space to previously tried parameter values.
Authors: We apologize for the insufficient detail. Failed runs are fed back by training an auxiliary failure-probability classifier on the same feature space; the multi-objective acquisition function is then augmented with a penalty term proportional to the predicted failure probability. This actively steers the search away from high-failure regions rather than discarding data or simply restricting the domain to previously sampled points. The revised Methods section will contain an explicit description of this procedure together with pseudocode. revision: yes
Circularity Check
No circularity: experimental MOBO+HITL workflow is self-contained
full rationale
The paper applies standard multi-objective Bayesian optimization to photonic curing parameters for metal-oxide capacitors, then augments it with a human-in-the-loop rule for failed runs. No equations, predictions, or first-principles results are derived; the claimed reduction in experimental rounds is an empirical outcome of running the optimizer on real fabrication data rather than a quantity forced by construction from fitted inputs or self-citations. The two objectives (capacitance-frequency dispersion and leakage current) are chosen by domain knowledge, not defined in terms of the optimization result itself. No load-bearing self-citation chain or ansatz smuggling appears in the provided text. The derivation chain is therefore independent of the target claim.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.