Accelerating computational materials discovery with artificial intelligence and cloud high-performance computing: from large-scale screening to experimental validation

1 Introduction

• AI models can predict quantum-mechanical results far faster and optimize structures, enabling exploration of chemical spaces larger than the set of known materials.The paper describes models as several orders of magnitude faster than direct simulation.Source: §1 · p. 2
• Recent studies generated large datasets of hypothetical materials without experimental validation, while technical validation of functional properties remained to be demonstrated.The paper notes that discovering technologically relevant materials is harder than finding synthesizable ones because they can be property outliers.Source: §1 · pp. 2–3
• Evaluating even readily generated candidate pools requires substantial computing resources, slowing progress despite faster ML-based potentials.Cloud HPC is presented as a way to evaluate more small-scale, complex calculation jobs and lower entry barriers through reproducible workflows.Source: §1 · p. 3
• The study combines AI and cloud HPC to identify and experimentally validate new solid-state-electrolyte compositions with targeted properties.The authors describe the process as rapidly identifying candidates and reducing time from concept to solution.Source: §1 · p. 3

2.1 Exploring a vast chemical space

• The search starts from 54 selected elements and deliberately avoids constraints based on ionic similarity, broadening the substitution space.Later-row transition metals, most lanthanides, noble gases, and actinides were excluded based on their stated relevance to battery materials or functional applications.Source: §2.1 · p. 3
• Ionic substitution generated 32,598,079 initial candidates from ordered ICSD structures using common oxidation states and structure prototypes.The approach used Pymatgen StructureMatcher and iso-valent substitutions.Source: §2.1 · p. 3
• The candidates cover 184 of 230 space groups and all seven crystal systems, expanding beyond matterverse.ai in prototype count, atom limit, and anion scope.Figure 1 presents element and space-group distributions for the initial candidates.Source: §2.1 · p. 4, Figure 1 · p. 4
• The broad substitution strategy is less thorough for any single composition than structure-prediction approaches such as AIRSS, USPEX, and CALYPSO, but is intended to sample material space more diversely.The authors say it has a less-biased focus on specific regions of material space.Source: §2.1 · p. 4
• The screening workflow concluded with 23 candidates, 18 of which had compositions not previously reported.Figure 2 depicts the main stages used to filter candidate materials.Source: Figure 2 · p. 5

2.2 Screening workflow for stable materials and solid electrolytes

• The two-stage workflow reduced 32.6 million candidates to 589,609 predicted stable materials, then applied electronic, electrochemical, and diffusivity filters to identify promising solid electrolytes.The first stage used ML potentials for relaxation and phase stability; later filters included band gap, redox potentials, and Li diffusivity.Source: §2.2 · pp. 5–6
• Practicality filters removed selected costly elements, materials with bulk or shear moduli above 30 GPa, and densities above 2.5 g cm−3, leaving 23 finalists, including 18 with previously unreported compositions.The workflow also evaluated ordered structures derived from partially substituted disordered Na-Li phases and kept the lowest-energy structures.Source: §2.2 · p. 6

2.3 Infrastructure for cloud-based materials discovery

• The cloud setup ran containerized ML and DFT jobs through SLURM on adjustable VM scale sets, with each job isolated in its own container.The VM scale set serves as the counterpart of a conventional HPC partition or queue.Source: §2.3 · p. 6, Figure 3 · p. 7
• A searchable CosmosDB database tracked job and chunk status to support restarts on pre-emptible or low-priority VMs; the study used approximately 1,000 VMs.The cloud environment supported tasks from Python scripting to DFT and molecular-dynamics calculations.Source: §2.3 · p. 6

2.4 Screening results and trends

• Approximately 1,000 VMs running for less than 80 hours produced ESW data for 8,795 materials and diffusivity data for 583 compositions and structures.The authors state that these datasets can guide future electrolyte design.Source: §2.4 · p. 7
• Early-row halides F, Cl, and Br tend to have larger electrochemical stability windows, oxides follow closely, and S- and Se-based chalcogenides have very small windows.The reported trend matched a previous study of Li-M-X ternary systems.Source: §2.4 · p. 8
• The 583 candidates passing ESW filters contained Br, Cl, F, or O; bromides and chlorides showed consistently good high-temperature Li diffusivities, while oxide and fluoride diffusivities varied widely.The authors associate the oxide and fluoride variation with sensitivity of conductivity to structural changes, citing the tetragonal and cubic phases of Li7La3Zr2O12.Source: §2.4 · p. 8

2.5 Final candidates

• Four rare-earth halides were selected for experimental validation: Li5YCl8, Li7Y2Cl13, Na2LiYCl6, and a Pna21 polymorph of Li3YCl6.The first three passed all property filters; the Li3YCl6 polymorph passed property-based filters.Source: §2.5 · p. 8
• Na2LiYCl6 was of particular interest because it showed Li–Na dual-conductivity properties.The study identifies a framework accommodating two distinct alkali cations as a potential route to broaden solid-electrolyte formulations.Source: §2.5 · pp. 8–9
• The selected halides had good DFT electrochemical stability windows, large DFT band gaps, and good predicted room-temperature Li conductivities from ab initio molecular dynamics.Na2LiYCl6 was singled out for its Li–Na dual-conductivity properties.Source: §2.5 · p. 8

2.6.1 Structure and composition

• PXRD showed the Na3YCl6 R-3 phase as the main phase for x = 1.5 to 3, while x = 0 to 0.5 matched trigonal Li3YCl6 P-3m1.Syntheses targeting Li3YCl6, Li5YCl8, and Li7Y2Cl13 yielded Li3YCl6 in its known P-3m1 phase; Na2LiYCl6 PXRD matched the predicted Na3YCl6 R-3 structure.Source: §2.6.1 · p. 9
• PXRD characterization was limited by humidity-affected crystallinity, Kapton-tape background, and broadened peaks that impeded accurate unit-cell refinement.Preliminary single-crystal XRD suggested successful synthesis of Na2LiYCl6 R-3; additional measurements were ongoing to confirm all series compositions.Source: §2.6.1 · p. 10
• XPS showed qualitative agreement between measured Li/Na ratios and nominal compositions, while SEM and EDS confirmed relative Na, Y, and Cl ratios in bulk samples.Source: §2.6.1 · p. 10

2.6.2 Conductivity

• At 100 °C, Na2LiYCl6 reached 2.2 × 10^-6 S/cm versus 6.0 × 10^-8 S/cm for Na3YCl6, while Li substitution lowered activation energy relative to the parent.Measured conductivity rose with increasing Li content; activation energies were 0.46–0.66 eV for x = 0.5 to 3 versus 0.82 eV for Na3YCl6.Source: Figure 5 · p. 9, §2.6.2 · p. 10
• The authors suggest enhanced conductivity and reduced activation energy in Li-substituted NaxLi3−xYCl6 may reflect Li+ transport and possible expanded diffusion channels or local disorder, pending validation of Na/Li pathways.Source: §2.6.2 · p. 10

Functional discovery

• With state-of-the-art AI and cloud HPC, researchers predicted half a million stable, potentially synthesizable materials from over 32 million candidates; filters found only a small number suited to targeted applications.Source: §3 · p. 11

Designing solid-state electrolytes

• The study predicted new Na- and Li-containing electrolyte compositions and experimentally validated NaxLi3−xYCl6-based compounds with conductivities considered viable for future electrolyte development.The workflow also recovered a decade of electrochemical stability-window trends, with halides showing the most promising balance of stability window and conductivity.Source: §3 · p. 11

Next steps

• The stringent electrochemical-stability-window filter excluded many candidates, including known sulfide Li10GePS12 and oxide Li7La3Zr2O12 electrolytes.Relaxing the filter may reveal additional electrolyte candidates among the 8,795 compounds that entered that filter.Source: §3 · p. 11
• The ML training set used only integer atomic occupancies, leaving disordered site occupancies and defects—common in fast Li conductors—outside the screened materials space.The authors call for extending the search beyond ordered, perfect crystals.Source: §3 · p. 12

Broader impact

• The integrated AI and cloud-HPC workflow is presented as an end-to-end approach with potential to accelerate identification, synthesis, and experimental validation of technologically relevant materials.Source: §3 · p. 12

4.1 AI model, training, and data

• The M3GNet potential was trained on curated Materials Project structures, with test errors of 29.9 meV/atom, 72.1 meV/Å, and 0.40 GPa.The dataset comprised 352,767 structures split into 317,489 training, 17,633 validation, and 17,645 test snapshots.Source: §4.1 · p. 12

4.2 DFT

• DFT and AIMD used VASP; trajectories covered at least 80 atoms, 200 ps, and 400–700 K, with Na2LiYCl6 simulated at 700–1000 K.The materials did not melt during the simulations.Source: §4.2 · pp. 12–13

4.3 Synthesis methods

• Stoichiometric NaCl, LiCl, and YCl3 mixtures were ground, pelletized, sealed in evacuated quartz ampoules, and annealed for five hours at 450–550 °C.Reagents and samples were handled in an Ar glovebox to minimize water contamination.Source: §4.3 · p. 13

4.4.1 XRD

• PXRD samples were covered with Kapton tape during measurement to limit moisture degradation.Measurements used a Rigaku SmartLab SE diffractometer with a Cu source.Source: §4.4.1 · p. 13

4.4.2 XPS

• Air-free XPS measurements used Li 1s, Na 2s, Y 3d, and Cl 2p spectra to calculate sample composition.Spectra were calibrated to the C 1s signal at 285 eV.Source: §4.4.2 · pp. 13–14

4.4.3 SEM and EDS

• SEM/EDS measured relative Na, Y, and Cl ratios at at least 10 sites per composition; Li was not detectable by EDS.Oxygen was observed at sample surfaces despite quick transfer from the glovebox.Source: §4.4.3 · p. 14

4.4.4 Electrochemical impedance spectra (EIS) measurement

• Ionic conductivity was determined from impedance measurements of pelletized electrolyte between stainless-steel electrodes across 25–100 °C.Cells were pressurized to 250 MPa, and rested for two hours at each temperature before measurement.Source: §4.4.4 · p. 14

• C.C., N.B., L.L., C.O., B.B., and M.T. work at Microsoft; two provisional patents were filed.Source: Competing interests · p. 15

• AI means artificial intelligence, while DFT and PBE mean density functional theory and Perdew–Burke–Ernzerhof, respectively.Source: Appendix A · p. 20
• AIMD denotes ab initio molecular dynamics, and EIS and ESW denote electrochemical impedance spectra and electrochemical stability window.Source: Appendix A · p. 20
• HPC, GPU, and ICSD denote high-performance computing, graphics processing unit, and Inorganic Crystal Structure Database.Source: Appendix A · p. 20
• PXRD, SEM, XPS, and EDS expand to powder X-ray diffraction, scanning electron microscopy, X-ray photoelectron spectroscopy, and energy-dispersive X-ray spectroscopy.Source: Appendix A · p. 20

• Supplementary plots show candidate electrochemical stability windows, PBE band structures, and AIMD Li and Na diffusivity results.Source: Appendix B · p. 20, Figure S1 · p. 21, Figure S2 · p. 22
• Table S1 reports SEM/EDS compositions normalized to Y, while Figures S5–S7 show PXRD patterns and Figure S8 shows elemental ratios.Source: Appendix B · p. 20, Table S1 · p. 21, Figure S8 · p. 25
• PXRD screenings for Li3YCl6, Li5YCl8, and Li7Y2Cl13 report trigonal Li3YCl6 and unreacted LiCl as the main observed phases.Source: Figure S5 · p. 23, Figure S6 · p. 24, Figure S7 · p. 24