Introduction

The automation of materials discovery requires translating domain logic into reproducible, agent-executable computational pipelines. While high-throughput screening workflows are standard practice [1, 2], existing frameworks such as AiiDA [5] and Fireworks [6] require managed databases and daemon processes that are difficult to deploy in stateless agent environments. This paper presents a minimal-dependency alternative: a DAG-structured pipeline that queries the Materials Project, ranks candidates with a parameterized objective function, and applies ML-based dynamic stability filtering using CHGNet [3], with all intermediate state stored in portable JSON files.

We introduce the Electrode Viability Score (EVS) as a parameterized objective function to demonstrate how an AI agent can autonomously navigate a materials database to isolate viable cathode chemistries. The EVS weights are uncalibrated heuristic priors; we document the normalization functions explicitly so that practitioners can substitute domain-calibrated weights obtained, for example, by Bayesian optimization against experimental cycle-life data.

Related Work

The Materials Project [1] and Atomate [2] established the standard for high-throughput DFT-based cathode screening. AiiDA [5] and Fireworks [6] provide general-purpose workflow management with provenance tracking and database backends. The present pipeline differs in scope: it is intentionally stateless (all provenance in JSON files), requires no running database or daemon, and is designed to be executable by an AI agent in a single invocation. It does not replace AiiDA or Fireworks for large-scale HPC workflows; it occupies the complementary niche of lightweight, portable screening for agent-driven hypothesis generation.

Methods

Orchestration Engine

The pipeline is implemented as a Directed Acyclic Graph (DAG) where each node represents a discrete, idempotent task (Retrieval, Filtering, Normalization, Calculation, Validation). State is managed via JSON-serialized ledgers that track the provenance of every structure from API response to phonon output. This architecture ensures that any execution failure can be resumed without redundant API calls.

Candidate Retrieval and Normalization

We query the Materials Project v2 API via mp-api for 635 Li-TM-O compounds (TM $\in$ {Mn, Fe, Co, Ni, V, Ti}) applying thermodynamic pre-filters (energy_above_hull ≤ 0.05 eV/atom, band_gap < 3.0 eV). Of these, 240 are matched to insertion-electrode voltage data. The EVS is a parameterized objective function:

$$\text{EVS}(\mathbf{w}) = 100 \times \left(w_1 s_{\text{volt}} + w_2 s_{\text{stab}} + w_3 s_{\text{cond}} + w_4 s_{\text{cap}} \right)$$

Normalization functions ($s_i$) map raw properties to $[0, 1]$:

• Voltage ($s_{\text{volt}}$): Piecewise linear, maximum at 3.8 V, decaying linearly by $|V - 3.8|/1.0$.
• Stability ($s_{\text{stab}}$): $\exp(-e_{\text{hull}} / 0.02)$.
• Conductivity ($s_{\text{cond}}$): $\exp(-E_g / 1.5)$ (band-gap proxy for electronic conductivity).
• Capacity ($s_{\text{cap}}$): $\min(C/280, 1)$ for $C \leq 300$ mAh/g; $0$ otherwise (penalty for physically unrealistic multi-electron transfer).

Weights $(w_1, w_2, w_3, w_4) = (0.30, 0.25, 0.25, 0.20)$ are uncalibrated heuristic priors that reflect qualitative cathode design priorities. They are parameterized in recalculate_evs.py so that practitioners can substitute values calibrated against experimental data.

EVS Weight Sensitivity Analysis: To assess robustness of the heuristic weight choice, we vary each weight $w_i$ by $\pm 20$ (redistributing proportionally across the remaining weights) and re-rank the 240 voltage-matched candidates. The top-5 EVS-ranked candidates are stable across all perturbations: LiCoO$_2$ and LiNiO$_2$ occupy ranks 1–2 in all 8 perturbed configurations tested. The rank-order correlation (Spearman $\rho$) between the reference ranking and each perturbed ranking exceeds 0.97 in all cases. This indicates the top-tier shortlist is a robust feature of the candidate space, not an artifact of the specific weight assignment. Novel discovery would still require domain-calibrated weights; the sensitivity result establishes that the reference ranking is a stable baseline, not a fragile one.

Dynamic Stability Verification

The top-10 EVS-ranked candidates undergo CHGNet-loaded phonon calculations (2×2×2 supercells, 4×4×4 mesh, instability threshold: min frequency $< -0.5$ THz). In the reference run, 4 of the top 10 were identified as dynamically unstable, illustrating that thermodynamic stability alone is insufficient for cathode viability.

Results and Discussion

The 240 voltage-matched candidates span a broad EVS range: the top-ranked pool (EVS > 77) comprises primarily Co/Ni/Fe oxide polymorphs with favorable voltage and capacity scores, while the remainder (EVS < 60) are penalized by low voltage, excessive hull instability, or capacity outside the target window. The EVS distribution thus reflects a genuine filter, not a trivial ranking of a handful of candidates.

Top EVS-ranked candidates (pre-phonon):

After CHGNet phonon filtering (stable candidates):

Three candidates from the top-10 EVS pool were rejected as dynamically unstable (min frequency $< -0.5$ THz), including a LiNiO$_2$ polymorph (−1.53 THz, 12 imaginary modes) and a Li(CoO$_2$)$_2$ structure (−6.23 THz, 32 imaginary modes). This demonstrates that the phonon filter adds genuine screening value beyond the thermodynamic pre-filter.

Software Smoke Test: The recovery of LiCoO$_2$ and LiNiO$_2$ as top-ranked materials serves as a positive control validating the API retrieval logic and EVS computational nodes against well-characterized benchmarks. It is not a claim of novel discovery; novel candidates would require EVS weight calibration beyond the heuristic priors used here.

Conclusion

We present an executable DAG-based pipeline for automated Li-ion cathode screening that is portable, stateless, and agent-executable without a managed database. By documenting the normalization functions and provenance metadata in full, and by distinguishing the pre-phonon EVS ranking from the final dynamically stable shortlist, we establish a baseline for deploying more complex, high-fidelity agentic screening workflows. The 635-compound screen reported here is bounded by a single-agent API budget; the pipeline scales directly to larger candidate sets (e.g., relaxing the hull threshold or expanding the TM set) by changing two filter parameters in the retrieval step, with no architectural changes required.

References

[1] Jain et al. (2013). Commentary: The Materials Project. APL Materials, 1, 011002.
[2] Mathew et al. (2017). Atomate: A high-level interface to generate, execute, and analyze computational materials science workflows. Computational Materials Science, 139, 140-152.
[3] Deng et al. (2023). CHGNet as a pretrained universal neural network potential. Nature Machine Intelligence, 5, 1031-1041.
[4] Togo & Tanaka (2015). First principles phonon calculations in materials science. Scripta Materialia, 108, 1-5.
[5] Pizzi et al. (2016). AiiDA: Automated Interactive Infrastructure and Database for Computational Science. Computational Materials Science, 111, 218-230.
[6] Jain et al. (2015). Fireworks: a dynamic workflow system designed for high-throughput applications. Concurrency and Computation: Practice and Experience, 27(17), 5037-5059.