acceptodds
Under review as a conference paper at ICLR 2027

BKBAgent: A Battery Parameter Identification Agent with Bayesian Experimental Design and Knowledge Graph

Abstract

Electrochemical battery models support battery design, state estimation, and control, but their accuracy depends on identifying internal parameters from operating measurements such as voltage. Existing work has applied large language model (LLM) agents to this task, using physical reasoning and simulation feedback to guide parameter updates. However, identifying coupled battery parameters from limited measurements is ill-posed: different parameter combinations can produce nearly identical responses. We introduce BKBAgent, a battery parameter identification agent that combines a physics-informed knowledge graph with Bayesian experimental design (BED). The agent maintains a joint belief over plausible parameter combinations and uses retrieved physical knowledge to propose candidates and anticipate their responses. Using these predictions, BED selects the next candidate by its expected change in the belief over parameter hypotheses. Simulator feedback updates the belief, and the agent retains parameter combinations whose simulated responses match the measurements. We evaluate BKBAgent on a benchmark comprising 108 test tasks with multi-parameter perturbations and large single-parameter shifts across operating rates. Under the same simulation budget, it reduces mean parameter identification error by 25.1% and 23.8% on the two task types, respectively, compared with an LLM-agent baseline. Ablations show that removing either retrieved physical knowledge or BED increases parameter identification error. These findings illustrate how structured domain knowledge and information-guided simulation can support scientific agents in reasoning over competing hypotheses and updating beliefs from feedback.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
LLM agents, AI for science, Bayesian experimental design, knowledge graphs, battery parameter identification, inverse problems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_AaGKOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

BKBAgent combines a physical knowledge graph with an LLM-maintained weighted set of complete battery-parameter hypotheses. The agent proposes parameter vectors and possible simulator outcomes, selects simulations using expected KL belief change, and updates its weights from actual feedback. This is computational experiment selection using fixed measurements, rather than acquisition of new physical measurements.

On 108 synthetic tasks derived from 18 independent perturbation parents, it improves parameter recovery relative to BSA and a closer GP-based reference. Additional evaluations cover ablations, three LLM backbones, unseen operating rates, and one measured cell. 57968_BKBAgent_A_Battery_Param

Critical strengths
  1. The scientific-agent application is supported by meaningful controlled comparisons. Methods share parameter domains, response-validation criteria, and a budget counting attempted solves, including initialization and failures. Point estimates are selected without access to parameter ground truth, and parent-level bootstrap respects dependence among rates, parameterizations, and seeds. Matched ablations distinguish external knowledge, relational retrieval, and selection policy. Relative to GP-BKB, the smaller improvements also have paired intervals favoring BKBAgent.

  2. The paper carefully distinguishes response matching, parameter recovery, and uncertainty representation. Outputs retain complete parameter vectors, and response validation includes timing and multiple response components. Held-out-rate predictions and a measured-cell case study complement synthetic parameter-recovery results. The authors also disclose limitations, including the absence of an exact Bayesian interpretation, increased API workload, and cases where improved point accuracy does not improve complete-vector recovery.

Critical weaknesses
  • The specific value of information-guided selection and the adequacy of the belief representation remain incompletely established.

    • The LLM supplies hypotheses, outcome probabilities, and conditional weights without a coherent joint probabilistic model. Equation (9) correctly shows that the selection utility includes belief drift as well as discrimination between possible outcomes. In Table 5, drift contributes about 31.6% of selected-candidate utility, and drift-only selection produces slightly greater reference improvement than the proposed utility, 0.0908 versus 0.0873.
    • These small, one-step diagnostics do not establish that drift-only selection is generally better. They do leave the proposed mechanism unresolved: end-to-end comparisons against uniform, discrimination-only, and drift-only selection are needed. Furthermore, a fixed support of 16 hypotheses cannot recover omitted regions, and the two-dimensional conditional diagnostics do not validate uncertainty representation in the full ten-dimensional problem.
  • Test-set independence is not fully verified, and the evaluation has limited independent breadth.

    • Appendix J.2 removes two Extreme parents whose complete parameter truths repeat development cases, but explicitly leaves possible endpoint duplication for coordinates 1 and 10 unverified. Complete separation should be checked and any affected results recomputed.
    • More broadly, the 108 tasks represent 18 independent parents under two parameterizations of the same cell family. The unseen-rate tests remain within the benchmark’s operating-rate range, while real-data validation covers one cell and subsequent discharges without parameter ground truth. These support a promising controlled result and predictive case study, but provide limited evidence of robustness across cell populations and model mismatch.
Questions for authors
  • Can you verify complete development/test truth separation, especially for Extreme coordinates 1 and 10, and report results after excluding any remaining duplicates?
  • How do uniform, discrimination-only, and drift-only selectors perform in full identification runs?
  • How sensitive are results to the initial hypothesis support and its size? Can the full-dimensional belief or its predictive uncertainty be assessed?
  • Does the advantage over GP-BKB persist under matched token budgets, alongside the existing matched simulation-budget comparison?
Reviewer_94XwOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Given observations of a battery behaviour: time series of voltage and current under different workloads and charging settings yobsy^{obs}, the goal is to infer the battery’s internal state θ\theta. We have access to a forward model from which data can be generated ysim=Ysim(θ)y_{sim} = Y^{sim}(\theta) and we aim to find argminthetad(Ysim(theta),Yobs)argmin_{theta} d(Y^{sim}(theta), Y^{obs}) for a given distance function d()d(). The authors also consider data from multiple “protocols”, each one producing its own time series and objective is a weighted sum of distances. If I understand correctly, this is the standard formulation of an inverse problems, we have data and a parameterized dynamics model, we just need the parameters.

In this case, the forward model is expensive, non-differentiable hence we need a method to find the parameters efficiently an dinteligently. One may use off-the shelf black box optimzers such as Bayesain optimization of CMA-ES. However these are general purpose algorithms. This paper considers an LLM based approach where the LLM is augmented with a physics based knowledge graph, extra battery specific information so that it can be prompted with past parameters and simulation outputs and withteh knowledge graph the LLM can respond with new candidate parameters to try. These candidate parameters are then passed through a Bayesian experimental Design step to find the single most informative parameter θ∗\theta^* to use for simulation. Expensive simulation is performed Ysim(θ∗)Y_{sim}(\theta^*), relevant histories and weights are updated and the method repeats.

Experiments are performed on an inverse problem with a range of ground truth parameter settings showing improvement over a recent agent, BSA (citation below), and black box optimization methods.

Critical strengths
  • builds upon the very recent Battery-Sim-Agent (BSA) that added an LLM into the loop for solving battery inverse problem that used to be solved by black box optimization methods.
  • this agent uses a persistent Knowledge Graph of battery information and therefore can make more informed decision than an agent who can only use in context current experiment data to make decision which appears to be the framework of BSA, this is a well motivated and simple idea.
Critical weaknesses

What is the machine learning problem being solved? Is this in the scope of ICLR?

The claim of the paper is that BKBagent can solve inverse problems for battery parameters better than prior methods and as I understand this is true. However

  • lack of ML Novelty the proposed solution, BKBAgent, does not appear to require new maths/statistics, network architectures, loss functions, optimization algorithms, data structures etc. I am not sure what new science I can learn from this paper.
  • battery specific audience The writing of the paper appears to assume the reader is already familiar with battery terminology. Technical terms are not explicitly described, the paper does not smoothly accommodate non-experts, instead non-experts are forced to infer them in real time or look them up elsewhere or ignore them and hope it won't matter. (in contrast Battery-Sim-Agent (BSA) Chen et.al. 2026 [1] does elegantly accommodate a general audience)
  • where the LLM is used is not very clear there are multiple contradictory ways the LLM is presented
    • Appendix Algorithm 1 refers to “LLM” two times
    • Appendix A Line 670 says there are 5 LLM calls
    • the main paper Figure 2 flow diagram has 3 blocks with the LLM (ChatGPT) symbol
    • the main body of the paper does seem to give descriptions of each step, the information is available to very determined readers
    • In contrast with BSA [1] whose algorithm 1 line 14 explicitly states calling LLM and Figure 1 shows a loop with one LLM block etc
  • hard to generalize as a corollary of the above points, it is not clear how this agent building recipe can be generalized to other applications.
  • no limitations I may have missed them, but in the paper and appendix I could not see a list of limitations.

The paper does build upon a recent agentic framework BSA who first (as I understand) proposed an agentic loop for this problem, and the improvement seems valid. However the apparent lack of ML novelty, the writing for a somewhat specific audience, the algorithm does not generalize, the meta-recipe is not clear, I struggle to recommend this paper for acceptance to ICLR in its current form.

[1] Battery-Sim-Agent: Leveraging LLM-Agent for Inverse Battery Parameter Estimation, Chen et.al, KDD 2026

Questions for authors

In BSA [1], the authors provide a table 1 that qualitatively contrasts their features of their agentic approach against standard black-box optimization. Can the authors please provide a similar table contrasting BKBAgent and BSA? What are the novel features and what capabilities do they unlock?

Reviewer_eoYpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces "BKBAgent": an LLM-agent-based system for battery parameter inference, motivated by downstream battery design, control and state estimation. The paper focuses on inferring model parameters from observed voltage/current measurements of a battery over time. The work is motivated by Bayesian Experimental Design (BED), here used to choose which simulation to run next, and utilises a knowledge-graph approach to inform the LLM's proposals.

BKBAgent is tested on battery simulation benchmarks with a reported ~25% improvement in point-estimate parameter error when compared with an alternative LLM-agent system, and a smaller (~6%) improvement over a GP-based version of the same pipeline.

Critical strengths

The paper uses a physics-based simulator with known ground-truth parameters as a benchmark, so parameter recovery can be measured directly and the evaluation is grounded in a level of realism needed for this task. This is complemented by a held-out prediction test on one real cell.

Critical weaknesses

Unclearly written. Some concepts in the pipeline are not formally introduced, and just appear in the text. An example of this is "archive", which appears (on page 6) and isn't well defined, nor linked to the set \hat{\theta} of section 3.1. Also, how some of the variables are estimated or calculated isn't clearly defined. For example, the probabilities that go into b and into equation 3, which are crucial for the Bayesian experimental design/decision framing, are only stated in passing in the main text, with the details in the appendix, and require some significant detective work to find, and the main text never states what these are actually probabilities of. This makes it difficult to assess the novelty of the methodology itself.

The impact of the work is not clear. If this is particularly impactful specifically to the battery design/inference domain, that is not clear. The gain over the closest baseline, GP-BKB, is only 0.006-0.007 NRMSE. Extreme Hit@5 is worse than GP-BKB, and the real-cell evidence is a single cell without parameter ground truth. Plus, specifically how impactful this method is outside of this domain is also unclear - the knowledge graph is battery-specific.

A significant weakness is that the probabilities in this task, which is motivated from a Bayesian design perspective, are generated by an LLM (b_t, p_m and the conditionals) in equation 3 all come from an LLM. This does not seem at all reliable. The authors acknowledge the resulting incoherence (equations 4 and 9) but do not really address it. There is a framework in probabilistic methods, e.g. simulation-based inference, to robustly estimate such probabilities. Whereas, here it is used only as an evaluation reference on 2-D slices. Furthermore, even if the probabilities are indeed correct, this is not well validated. This aspect is crucial to the framing of the paper.

The headline metrics are parameter-based, so less targeted at the full task that motivates the work. Plus these parameter-based metrics are the root-mean-square error against a single ground truth \theta^*, which, if we interpret it from a Bayesian standpoint, can at best assess a single point-wise summary of the posterior (the posterior mean) and is blind to its coverage. Moreover, the reported point estimate is the best-fit evaluated vector rather than any functional of the belief. This is a significantly smaller scope than the motivation for the paper.

Questions for authors

Is it correct that the probabilities that enter into equation 3 are themselves generated by an LLM, and, if so, what makes you trust these are actually probabilities?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.