El Agente Potente: High-Throughput Agentic Atomistic Simulations

1 Introduction

• High-throughput MLIP studies must coordinate structure preparation, relaxation, sampling, property calculations, validation, failure handling, and data aggregation.Workflows can differ substantially between scientific problems.Source: §1 · p. 2
• Existing frameworks such as AFLOW (9), AiiDA (10; 11), Pyiron (12; 13), and Atomate2 (14) automate simulations but require substantial package and domain expertise.Source: §1 · p. 2
• Prior El Agente systems explored autonomy through multi-agent reasoning, typed graphs and coding, tool generation, and agents embedded within structured graphs.These approaches span flexibility in reasoning, workflow construction, and tool construction while seeking reliable scientific execution.Source: §1 · p. 2
• Potente combines typed execution graphs for standardized workflows with coding mode for customized procedural logic, reusing supported functions where possible.Graphs provide explicit control over data flow, validation, branching, provenance, and concurrent execution.Source: §1 · p. 2
• Its evaluation includes three systematic benchmarks and seven case studies spanning materials discovery, molecular energy landscapes, adsorption, and catalytic pathways.The benchmarks examine reproducibility, foundational MLIP performance, and LLM token cost, including comparisons with open-ended skill-based agents.Source: §1 · p. 2

2.1 Architecture of El Agente Potente

• The architecture routes natural-language requests to typed graph workflows for standardized campaigns or agent-generated Python for customized control flow, while deterministic functions perform validated calculations.Graph nodes support structure preparation, relaxation, optional filtering, property calculations, and aggregation; mapping, streaming, and joins coordinate execution, with outputs, failures, decisions, artifacts, and provenance retained as workflow state.Source: Figure 1 · p. 3, §2.1 · p. 4

2.2.1 Reproducibility of high-throughput calculations

• Repeated Potente graph executions matched scripted outputs across bulk and shear moduli and heat capacity, while retaining typed failures rather than interrupting campaigns.Across five executions, both repeated-run standard deviations and parity comparisons were reported as narrow or essentially along y = x; the sole occasional failure occurred during tolerance-sensitive Niggli reduction.Source: §2.2.1 · pp. 4, 6

2.2.2 Agentic benchmarking of foundational machine learning interatomic potentials

• Bulk-modulus MAE favored MatterSim (35), while shear-modulus MAE favored Orb-OMAT (36) and shear-modulus R2 favored MatterSim (35), showing property-dependent rankings.For bulk modulus, MAE/R2 were MatterSim 13.99 GPa/0.873, MACE-OMAT (33) 15.69 GPa/0.871, and Orb-OMAT (36) 16.88 GPa/0.863; for shear modulus, MAE/R2 were Orb-OMAT 13.94 GPa/0.624, MatterSim 14.05 GPa/0.663, and MACE-OMAT 15.24 GPa/0.439. Potente's coding mode built the comparative workflow using existing validated functions, and all 3,000 model–structure calculations completed.Source: §2.2.2 · p. 6

2.2.3 Comparison with a skill-based coding agent

• Both systems completed the HKUST-1 CO2 GCMC task, but Potente's graph offered reproducible estimates and stable resource use while Claude Code constructed workflows anew across runs.Only two of five Claude Code runs considered framework pre-relaxation; because backends and sampling protocols differed, the comparison assesses workflow behavior, not adsorption accuracy, and the short simulations support screening-level predictions.Source: §2.2.3 · p. 7, Table 1 · p. 7

2.3 Case studies

• Seven tailored research problems demonstrate Potente's ability to execute atomistic simulations across diverse application scenarios.The case studies include materials discovery and molecular energy-landscape exploration.Source: §2.3 · p. 7

2.3.1 Materials Discovery

• Five selected lead-free halide perovskites lay 0.0151–0.0579 eV atom−1 above the MLIP convex hull, with predicted band gaps of 0.774–1.609 eV.CsSnCl3 was the most stable at 0.0151 eV atom−1 above hull with a predicted 1.312 eV band gap; the results came from literature-guided proposal, generated structures, MLIP relaxation and down-selection, and band-gap prediction, and are preliminary screening rather than final validation.Source: §2.3.1 · p. 9
• Selected lithium-conductor candidates spanned 0.0618–0.1038 eV atom−1 above hull and predicted 300 K conductivities of 5.89–65.44 mS cm−1, with tellurides higher-conductivity but more metastable.Li12Sb4Se16 was 0.0618 eV atom−1 above hull with 6.56 mS cm−1 conductivity; Li12Sb4Te16 reached 65.44 mS cm−1 at 0.1038 eV atom−1. The predicted conductivities are screening-level indicators given short MD trajectories, finite cells, and model limitations.Source: §2.3.1 · pp. 9–10

2.3.2 Molecular Energy Landscape Exploration

• For tetraethylene glycol dibenzoate, MLIP conformer search found 84 local minima across about 0.535 eV, and DFT reranking retained the MLIP-lowest conformer as the lowest-energy structure.The ten lowest MLIP conformers spanned 0.129 eV and the DFT-ranked ten spanned about 0.135 eV; the lowest conformer was optimized and used for vibrational analysis. The subsection also describes agent-selected collective variables and metadynamics analysis of alanine dipeptide's free-energy surface.Source: §2.3.2 · pp. 10, 12, Figure 4 · p. 11
• In 50 ps, metadynamics visited 2,429 of 5,184 bins (~47%) and sampled C7eq-like and C5-like basins from C7ax, reproducing qualitative landscape topology but not quantitative relative energies.The lowest sampled region was around (ϕ, ψ) = (−77.5°, 102.5°), associated with C7eq, and a second low-energy region around (−147.5°, 157.5°) was associated with C5. The workflow also generated broadened IR and Raman spectra and machine-readable peak tables after MLFF-based exploration and targeted higher-level refinement.Source: §2.3.2 · pp. 12–13
• Given an underspecified alanine-dipeptide request, the agent inferred periodic Ramachandran torsions as collective variables and configured two-dimensional well-tempered metadynamics.It derived the torsions from molecular connectivity, selected 5° bins, 1.0 kJ mol−1 Gaussian hills with 10° widths, deposition every 50 fs, and a bias factor of 10; the Orb-OMOL backend performed the numerical dynamics and bias-force calculations. The workflow preserved its configuration, script, trajectory, free-energy grid, and artifact summaries.Source: §2.3.2 · pp. 12–13

2.3.3 Adsorption and Catalytic Reaction Workflows

• MACE-OC20’s lowest H2/Fe(100) configuration was dissociative, at approximately −2.47 eV per H2; the runner-up differed by 2 meV, and the next minimum was 0.25 eV higher.H–H dissociation occurred during relaxation despite molecular H2 starting structures. The adsorption energy is an MLIP prediction, not quantitatively validated against spin-polarized DFT or experiment; higher-level validation is needed.Source: §2.3.3 · pp. 15–16
• For H2 adsorption, the workflow constructed and symmetry-reduced candidate sites, sampled orientations and anchoring choices, and allowed relaxation to find dissociated states rather than enforcing intact H2.The workflow translated a request for lowest-energy configurations into executable, provenance-preserving exploration using MACE-OC20 while retaining the numerical energy and force evaluations in the MLIP backend.Source: §2.3.3 · pp. 15–16
• MACE-OC20 predicted a 1.09 eV activation barrier and +0.53 eV reaction energy versus DFT’s 1.41 eV and +0.85 eV; MACE-OMAT predicted 1.79 eV and −2.28 eV.MACE-OC20 reproduced the endothermic character and approximate DFT energetics, whereas MACE-OMAT predicted an exothermic reaction. The same endpoint structures and search protocol exposed substantial model dependence.Source: §2.3.3 · p. 16
• RIPPEN had the highest log10 SCO2/H2O selectivity (3.07), followed by PEGBUA (2.52) and GAVYAG (2.36); GELVID01 was nearly non-selective (0.04), whereas ARAHIM was water-selective (−0.37).RIPPEN and PEGBUA best met the combined criterion of appreciable CO2 affinity and suppressed H2O uptake; high CO2 heat of adsorption alone was insufficient. CO2 Henry coefficients were lowest for ARAHIM and highest for PEGBUA (17.82), while average CO2–framework interaction energies ranged from −0.438 to −0.631 eV.Source: §2.3.3 · pp. 17–18

3 Discussion

• Typed graphs control scientific state, validation, branching, failure handling, provenance, and parallel execution; coding mode supplies task-specific logic when graph components are insufficient.The coding agent reuses existing Potente functions for supported calculations, leaving numerical execution to validated components while LLM reasoning organizes scientific actions.Source: §3 · p. 18
• Execution-graph workflows reproduced manually scripted numerical behavior with provenance and failure handling, while coding mode built customized comparisons using the same validated functions.Against open-ended coding, established graphs can reduce repeated runtime reasoning, while open-ended coding remains useful for creating new procedures.Source: §3 · p. 18
• Prediction accuracy remains limited by model transferability and numerical choices, and structured execution cannot ensure scientifically optimal workflows or parameters.Higher-fidelity calculations and domain-specific validation remain necessary when quantitative accuracy is required.Source: §3 · p. 18
• The paper identifies adaptive workflow selection and computational fidelity as next steps, with uncertainty, model disagreement, failed validation, or setting sensitivity as possible triggers for higher-fidelity calculations.Those calculations could refine screening decisions or provide data to improve models; experimental automation could support closed-loop, provenance-aware workflows across fidelity levels.Source: §3 · p. 18

4.1 Computational materials science and chemistry software

• The project used uv to manage a shared Python 3.12 environment, with ChatGPT 5.5 as the primary LLM and ChatGPT 4.1 powering routing agents.The shared environment included computational packages for single-agent execution; selected packages were forked where needed to resolve dependencies or expose programmatic interfaces.Source: §4.1 · p. 19
• ASE provided the common atomic-structure representation, geometry relaxations, molecular dynamics, and the MLIP calculator interface for energies, forces, and stresses.These calculator outputs were used in downstream property calculations.Source: §4.1 · p. 19
• Pymatgen queried Materials Project databases through mp-api, supported MLIP-energy phase-diagram construction, and converted structures between ASE and pymatgen objects.Source: §4.1 · p. 19
• Property calculations used MACE, MatterSim, Orb, and PET MLIP backends, with MatCalc connecting MLIP evaluators and structure representations to atomistic workflows.Elastic properties used stress–strain calculations and pymatgen-based elasticity analysis; lattice-dynamics workflows used phonopy, phono3py, and symfc.Source: §4.1 · p. 19
• Torch-dftd supplied the MACE-DAC-1 dispersion correction, while py3Dmol, matplotlib, and plotly supported molecular and crystal visualization and post-processing.Source: §4.1 · p. 19

Data availability

• The project repository provides benchmark results, generated user prompts, workflow outputs, and, where applicable, input structures, generated scripts, and post-processing outputs.Source: Data availability · p. 19

Code availability

• Source-code repositories are listed for the main packages, including ASE, pymatgen, MACE, MatterSim, Orb, MatCalc, torch-dftd, UPET, phonopy, and phono3py.Source: Code availability · p. 20
• A complete package list and environment specification will be provided with the repository, and third-party packages, pretrained MLIP models, and external databases were used under their respective licenses.Source: Code availability · p. 20

A.1 Graph inputs and shared state

• A typed workflow request supplies structure sources, model and calculation settings, and optional filtering or knowledge-graph options; deterministic Python nodes execute the graph.The outer agent converts the user request into the typed invocation, while an inner LLM router selects downstream property nodes.Source: §A.1 · p. 26, Execution graph implementation · p. 26

A.2 Candidate and reference streams

• Candidate structures are streamed and relaxed individually, while an optional Materials Project branch relaxes reference entries and builds an MLIP-consistent phase diagram.The reference branch is active for generated-composition inputs when phase-stability filtering is enabled; otherwise it supplies a placeholder and filtering passes structures through.Source: §A.2 · p. 26

A.3 Join synchronization and phase-stability filtering

• A tagged join waits for both candidate outcomes and the phase-diagram token before filtering; candidates above the configured hull cutoff skip property calculations.The filter can retain only the lowest-energy-above-hull candidates per reduced composition, and relaxation failures remain typed outcomes.Source: §A.3 · pp. 26, 28

A.4 Property routing and decision dispatch

• Successful structures go to an LLM router that returns validated typed property-node configurations, while failed, filtered, or relaxation-only cases receive terminal results.Dispatch targets include molecular dynamics, ionic conductivity, elasticity, phonon, density-of-states, Widom insertion, and GCMC calculations.Source: §A.4 · p. 28

A.5 Aggregation, artifacts, and knowledge-graph persistence

• Property and terminal results are grouped by structure, assembled into per-structure outputs, and written as deterministic artifacts; selected semantic results can be persisted after completion.The knowledge graph stores workflow, relaxation, routing, phase-diagram, and typed-property records, not runtime streams or calculator objects.Source: §A.5 · p. 28

A.6 Parallelism model

• The graph exposes parallel work across candidate relaxations, reference relaxations, and routed property calculations, while runtime limits and safety guards bound hardware concurrency.Candidate and reference branches can run concurrently before synchronization.Source: §A.6 · p. 28

B.1 Phase stability analyses for materials discovery

• Convex-hull diagrams show the lowest-Ehull generated candidates for five lead-free halide perovskites and five Li-ion conductors against support entries and MLIP hull tie lines.Generated candidates are marked by green squares; support entries are open circles.Source: Figure S2 · p. 30, Figure S3 · p. 31

B.2 Molecular case studies

• Molecular case studies include broadened DFT-reranked conformer IR/Raman spectra and alanine-dipeptide metadynamics mapping backbone-dihedral free energy and sampling at 300 K.The free-energy minimum is marked near (ϕ, ψ) = (−77.5°, 102.5°).Source: Figure S4 · p. 32, Figure S5 · p. 32

• The Figure S6 scatter plot relates CO2 isosteric heat of adsorption to CO2/H2O Henry selectivity for five representative MOFs; green points agree well with reference values, while red points show modest deviations.Source: §B.3 · p. 33, Figure S6 · p. 33
• Point color distinguishes predicted adsorption properties that agree well with reference values from those showing modest deviations.Source: §B.3 · p. 33

• The benchmark prompt asks for MatterSim, Orb-OMAT, and MACE-OMAT predictions for each structure and CSV outputs containing bulk and shear moduli.Source: §B.4 · p. 33, Figure S7 · p. 34
• The requested comparison plots the calculated moduli against ground truth in elasticity_1000.csv, with units in GPa.Source: §B.4 · p. 33
• Parity plots compare MLIP-predicted and DFT-reference Voigt–Reuss–Hill bulk and shear moduli, with the dashed line marking ideal agreement.Source: Figure S7 · p. 34

Potente and Claude Code comparison

• Across five Potente runs, HKUST-1 CO2 uptake averaged 73.53 ± 3.30 g/L, or 4.60 ± 0.21 molecules per cell, using MACE-DAC-1.Average interaction energy was −1.09 ± 0.04 eV, and average interaction energy per adsorbate was −0.247 ± 0.002 eV/CO2.Source: §B.5 · p. 34
• Potente’s 5,000 production steps yielded screening-level adsorption predictions, not fully converged adsorption isotherm data.Source: §B.5 · p. 34
• Five AtomisticSkills runs averaged 1.25 ± 0.57 mmol/g CO2 loading and 13.81 ± 0.77 kJ/mol estimated heat of adsorption using MACE-MH-1.The reported 8–9% loading variation was consistent with expected stochastic variation from short production windows.Source: §B.5 · pp. 34–35
• The authors attribute the discrepancy between the two agents’ adsorption estimates to their use of different MACE models.Source: §B.5 · pp. 34–35

B.6 Reproducibility benchmarks

• Standard-deviation distributions across five trials are shown for bulk modulus, shear modulus, and harmonic heat capacity at 300 K, with median and 95th-percentile markers.The figure describes the distribution of standard deviations across independent trials.Source: Figure S8 · p. 36, §B.6 · p. 36
• Each parity-plot point gives one material’s five-run mean for each workflow; the dashed line marks perfect agreement (y = x).The properties shown are the Voigt–Reuss–Hill bulk and shear moduli and harmonic heat capacity at 300 K.Source: Figure S10 · p. 37

C.1.1 Perovskite discovery

• The proposals cover Pb-free K/Rb/Cs halides with Ge or Sn, and are framed as reasonable underexplored candidates rather than proven novel compositions.The screening prompt requested five novel and reasonable ABX3 compositions.Source: §C.1.1 · pp. 37–38
• Each of five proposed compositions was screened by generating ten structures, relaxing them with MACE-OMAT, selecting the lowest-hull candidate, and predicting its density-of-states band gap.All five workflow calls completed successfully.Source: §C.1.1 · p. 38
• CsSnCl3 had the lowest selected energy above hull (~0.015 eV/atom), while RbGeI3 and KGeBr3 were highlighted as balanced candidates with ~1.5–1.6 eV gaps.The shortlist interpretation identifies RbGeI3 and KGeBr3 as the most promising absorber-oriented candidates in this screen.Source: §C.1.1 · p. 38
• RbSnBr3 showed predicted metastability but a smaller computed gap, whereas KGeI3 had a ~1.28 eV gap and the highest energy above hull among the five selected structures.These comparisons inform the shortlist interpretation rather than establish experimental performance.Source: §C.1.1 · p. 38
• The reported results are MLFF/DOS screening predictions; the stated next validation step is higher-level DFT relaxation and band-structure calculation, ideally including spin–orbit coupling for Sn/I/Br systems.The passage presents these as recommended follow-up calculations.Source: §C.1.1 · p. 38

C.1.2 Superionic lithium-ion conductor discovery

• The proposals extend heavier, softer selenide and telluride Li3MX4 analogues as exploratory superionic-conductor candidates.The passage describes Li12M4X16 as a four-formula-unit cell of Li3MX4.Source: §C.1.2 · p. 39
• The workflow generated ten structures per composition, relaxed them with MACE-OMAT, ranked by energy above hull, and ran 300 K conductivity MD on the lowest-hull structure.The protocol used a 2 × 2 × 2 supercell with 100 ps equilibration and 100 ps production; all five workflow calls completed successfully.Source: §C.1.2 · pp. 39–40
• Li12Bi4Se16 was also near-hull, but its Bi(V)-selenide chemistry was flagged for caution despite the reasonable MLFF screening result.Source: §C.1.2 · p. 40
• Li12As4Te16 and Li12Sb4Te16 had predicted conductivities of 43.2 and 65.4 mS/cm, respectively, but hull energies at or above ~0.10 eV/atom, especially for Sb–Te.The tellurides are described as high-risk/high-reward leads.Source: §C.1.2 · p. 40
• Li12Sb4Se16 and Li12As4Se16 were prioritized as balanced candidates, combining near-hull stability with predicted 300 K conductivities of 6.56 and 9.14 mS/cm, respectively.Li12Sb4Se16 had the lowest energy above hull in the screen; Li12As4Se16 had the second-lowest hull energy and slightly higher conductivity.Source: §C.1.2 · pp. 39–40

C.2.1 Low-energy conformer search and PySCF validation

• 84 neutral-singlet C22H26O7 conformers were returned by the Orb-OMOL multiple-minimum Monte Carlo search, with minimum_001.xyz lowest at −37567.706940 eV.These are MLFF potential energies, not DFT electronic energies or thermochemical free energies; PySCF single-point calculations reranked the ten lowest conformers.Source: §C.2.1 · pp. 41–42
• The PySCF-optimized conformer had no significant imaginary frequencies at the 50 cm−1 threshold, although three floppy modes occurred at −43.94, −28.29, and −8.36 cm−1.PySCF optimization and IR and Raman spectral analysis used wB97X-D4/def2-TZVP; the optimized electronic energy was −1380.740477950003 Hartree.Source: §C.2.1 · pp. 42–43

C.2.2 Two-dimensional enhanced sampling of alanine dipeptide

• 0.000 kJ/mol was the minimum free energy at (phi, psi) = (−77.5°, 102.5°), with 2429/5184 bins visited across 100001 samples.The calculation used two-dimensional well-tempered metadynamics at 300 K, with a 0.5 fs timestep and 100,000 production steps.Source: §C.2.2 · pp. 44–45

C.3.1 Adsorption-site search for H2 on Fe(100)

• −2.5177812576 eV was the best H2/Fe(100) adsorption energy; the search retained 10 distinct relaxed minima from 12 site/orientation candidates.The workflow used MACE-OC20.Source: §C.3.1 · p. 47
• Potente constructed and validated a 160-atom bcc Fe(100) slab, ran a MACE-OC20 H2 adsorption-site search, ranked relaxed candidates, and saved structures, trajectories, and results.The retained relaxed minima were saved as minimum_01.extxyz through minimum_10.extxyz.Source: §C.3.1 · pp. 47–48

C.3.2 Transition-state search on Fe(110)

• MACE-OC20 predicted a 1.095 eV barrier and +0.528 eV reaction energy, versus MACE-OMAT’s 1.789 eV barrier and -2.278 eV reaction energy.The comparison used seven-image pathways with the bottom two Fe layers fixed; the cell was kept fixed.Source: §C.3.2 · pp. 48–49

C.4.1 CO2 Widom insertion

• PEGBUA_clean had the highest reported CO2 Henry coefficient among the five MOFs: 17.820547 mol/kg Pa.The MACE-DAC workflow relaxed each framework before insertion at 300 K; its default two-fold protocol performed 10,000 trial insertions per fold, 20,000 total attempts per MOF–adsorbate pair.Source: §C.4.1 · pp. 49–50

C.4.2 H2O Widom insertion

• For H2O insertion, C42H24Ni6O30 had a Henry coefficient of 1.659492, average interaction energy of -0.618802, and heat of adsorption of 62.199618.The MACE-DAC workflow relaxed five MOF CIF structures before insertion at 300 K; the default protocol used 10,000 insertions per fold, 20,000 total attempts per pair.Source: §C.4.2 · p. 50

Benchmark specification

• The requested benchmark covers bulk_modulus and shear_modulus predictions from MatterSim, Orb-OMAT, and MACE-OMAT for structures in the CIF dataset.The ground-truth comparison uses elasticity_1000.csv.Source: §C.5.1 · p. 51

Workflow execution

• The proposed script uses each CIF filename stem as mp_id, runs the MLFF workflow for each structure and model, and extracts both elastic moduli.It requests relaxed elastic-property calculations through the workflow router.Source: §C.5.1 · pp. 51, 55

Outputs and comparison

• The script saves one CSV per model and generates parity plots comparing predictions with ground-truth values.It is designed to detect the ground-truth identifier and modulus columns and search workflow outputs for common modulus field names.Source: §C.5.1 · pp. 51, 53, 54

C.5.1 Elasticity benchmark

• 1 eV/Å3 equals 160.21766208 GPa; the script converted predictions to GPa while preserving raw eV/Å3 values.Potente used a direct script with get_mlff_calculator, relax_structure, and run_elasticity for MatterSim, Orb-OMAT, and MACE-OMAT.Source: §C.5.1 · pp. 61, 63
• 13.99 GPa was MatterSim’s bulk MAE after unit correction.MatterSim’s bulk RMSE was 27.49 GPa and R2 was 0.873; its shear MAE was 14.05 GPa, RMSE was 27.13 GPa, and R2 was 0.663.Source: §C.5.1 · p. 63
• 16.88 GPa was Orb-OMAT’s bulk MAE after unit correction.Orb-OMAT’s bulk RMSE was 28.50 GPa and R2 was 0.863; its shear MAE was 13.94 GPa, RMSE was 28.67 GPa, and R2 was 0.624.Source: §C.5.1 · p. 63
• 15.69 GPa was MACE-OMAT’s bulk MAE after unit correction.MACE-OMAT’s bulk RMSE was 27.70 GPa and R2 was 0.871; its shear MAE was 15.24 GPa, RMSE was 34.99 GPa, and R2 was 0.439.Source: §C.5.1 · p. 63
• The default parity plot was replaced with the GPa-corrected version, and the default metrics file was updated accordingly.The GPa-specific plot copy remained available, and the original unconverted plot was backed up.Source: §C.5.1 · p. 64