High-Throughput Energy Screening

• High-throughput energy screening (HTES) is a systematic computational approach that integrates automation and cost-effective descriptors to rapidly evaluate large chemical spaces for energy applications.
• It employs a multi-stage workflow—from candidate library generation and low-cost prescreening to automated property prediction and data mining—to efficiently prioritize promising materials.
• HTES has transformed materials discovery in fields like catalysis, thermoelectrics, photovoltaics, and batteries by significantly reducing computational costs and accelerating experimental validation.

High-throughput energy screening (HTES) is the ensemble of computational and/or data-driven methodologies designed to rapidly and systematically evaluate large libraries of materials or molecular structures for targeted energy-related properties. This paradigm has become a cornerstone of rational energy materials discovery—spanning catalysis, thermoelectrics, photovoltaics, batteries, superconductors, and beyond—by integrating automation, cheap descriptors, surrogate models, and scalable first-principles calculations to triage vast chemical spaces and funnel resources toward experimentally actionable leads.

1. Core Principles and HTES Workflow Architectures

The canonical HTES pipeline is structured as an automated, multi-stage process. Its modularity and scalability are essential to match the combinatorial complexity of modern materials and formulation spaces (Afzal et al., 2019, Wang et al., 2017, John et al., 2018, Liu et al., 2022):

1. Model Specification and Benchmarking: Choice and validation of a predictive model (e.g., DFT with functionals, cluster expansion, ML surrogate).
2. Candidate Library Generation: Combinatorial enumeration, substitution matrices, or data-mined structural motifs are used to generate tens-of-thousands to millions of hypothetical structures (Stafford et al., 2023).
3. Low-cost Prescreening: Application of property heuristics (formation energy, electronic or geometric proxies, descriptor thresholds) rapidly eliminates non-candidates without expensive calculations.
4. Automated Property Calculation or Prediction: Batched high-fidelity calculations (DFT total energies, adsorption energies, band gaps, vibrational/phonon, or transport properties) are managed by scripts or workflow managers, with error handling and data tracking (Mathew et al., 2016).
5. Automated Data Mining and Filtering: Threshold-based selection, ranking, and clustering (e.g., overpotentials, zT, SLME, redox window) are applied prior to high-level theory or experimental prioritization.
6. Iterative Validation and Feedback: Experimental feedback or model recalibration (with, e.g., Bayesian or uncertainty-aware corrections (Pyzer-Knapp et al., 2015)) generates new predictive cycles, supporting adaptive screening.

This architecture is implemented in domain-specific tools (CE Screen (Wang et al., 2017), MPInterfaces (Mathew et al., 2016), SeA (Ko et al., 2022)), and is the computational substrate beneath all leading HTES campaigns.

2. Descriptor Engineering and Physics-Driven Surrogates

Accelerating HTES demands surrogate models or descriptors that map the many-body problem to computationally cheap but predictive metrics:

• Catalysis: Adsorption energy $E_{\text{ads}}$ of key intermediates is the principal descriptor ($$E_{\text{ads}} = E_{\text{slab+ads}} - E_{\text{slab}} - E_{\text{adsorbate}}$$), enabling activity trends via volcano plots (Afzal et al., 2019, Huo et al., 17 Dec 2025). Linear scaling relations between intermediates lead to dimensionality reduction and efficient screening.
• Thermoelectrics: Various descriptors are tailored to capture electronic and lattice performance, including:
  • The "electronic fitness function" $t = (\sigma/\tau)\,S^2 / N^{2/3}$ encapsulates decoupling between $\sigma$ and $S$, identifying optimal band-structure complexity (Xing et al., 2017).
  • For rapid ranking without BTE, $\mu \propto \epsilon_0^2 / m^*$ and $PF \propto \epsilon_0^2 / m^*$, with $\epsilon_0$ (dielectric constant) and $m^*$ (carrier mass) from simple DFT (Deng et al., 2021).
  • In binary chalcogenides, the composite $$\chi = (m^*_{d})^{3/2} \hbar k_B \rho v_l^2 / (m_c^*)^{5/2} E_d^2$$ serves as a proxy for $PF_{max}$, while the Grüneisen parameter $\gamma$ (from elastic constants) proxies for lattice anharmonicity and low thermal conductivity (Jia et al., 2019).
• Photovoltaics: Band gap ($E_g$), band-edge positions (NHE-aligned), and derived figures (e.g., SLME for absorbance/thickness effects) encode optoelectronic suitability (Liu et al., 2022, Stafford et al., 2023, Sahni et al., 2019).
• Superconductors: The fast EPC descriptor at the $\Gamma$-point, $$\lambda_{\Gamma} = \sum_\nu (\tilde\omega_{\Gamma\nu}^2 - \omega_{\Gamma\nu}^2) / 4\omega_{\Gamma\nu}^2$$, permits rapid pre-selection before full Brillouin-zone calculations (Wang et al., 2022).
• Batteries: Tracer diffusivity ($D$) from pinball or BOMD models allows direct ranking of Li-ion conductors (Kahle et al., 2019).
• ML-Empowered Surrogacy: Graph neural networks (GNNs), message-passing NNs, and Gaussian-process calibration models deliver sub-meV or sub-eV errors for molecular and periodic property prediction, enabling screening at scales beyond DFT (John et al., 2018, Pyzer-Knapp et al., 2015, Huo et al., 17 Dec 2025).

3. Integration of Automated Workflows and Data Infrastructure

Large-scale HTES is sustained by automation frameworks that orchestrate structure generation, property calculation, data handling, and error correction:

• CE Screen: Cluster expansion on MatCloud platform for doped/disordered alloys with systematic selection/DFT of training structures, predictive cross-validation, and auto-reporting (Wang et al., 2017).
• MPInterfaces: Automated slab/surface/interface builder (leveraging pymatgen and CatKit), robust VASP/LAMMPS pipelines with checkpointed error handling, implicit solvent corrections, and built-in analysis (surface energies, Wulff shapes) (Mathew et al., 2016).
• SeA: Black-box, linear-scaling hybrid DFT (SCDM + exx + ACE) with non-iterative orbital localization, low-rank exchange compression, and O($N$) scaling for large, finite-gap systems. Automation includes array job submission, parallelization, and fail-safe restarts, tailored for large molecular and condensed-phase datasets (Ko et al., 2022).
• Active Learning and ML-Aided Surrogates: Automated retraining and data acquisition (e.g., via active learning for DNN potentials), hybrid chemical/ML outlier detection, and tight integration with experimental or simulated data streams enable truly adaptive screening (Huo et al., 17 Dec 2025, Hußner et al., 2023).
• Practical Throughput: Well-designed pipelines can batch and process tens-of-thousands to millions of structures per day for surrogate/ML predictions and hundreds to thousands for first-principles workflows, with computational error rates and human intervention minimized by built-in validation steps (Afzal et al., 2019, Ko et al., 2022, Huo et al., 17 Dec 2025).

4. Case Studies Across Energy Materials Domains

HTES methodologies are instantiated and validated across numerous scientific domains:

Application	Descriptor / Metric	Screening Scale	Representative Papers
Electrocatalysis	$E_{\text{ads}}$, $\eta_{\text{lim}}$	750–15,911	(Afzal et al., 2019, Huo et al., 17 Dec 2025)
Thermoelectrics	$t$, $\chi$, $\gamma$, $\mu$, $PF$	75–243	(Xing et al., 2017, Deng et al., 2021, Jia et al., 2019)
Batteries	$D$ (tracer), $E_a$	~1,300	(Kahle et al., 2019)
Photovoltaics	$E_g$, SLME, redox windows, PCE	10^3–10^6	(Stafford et al., 2023, Liu et al., 2022, Sahni et al., 2019)
Superconductors	$\lambda_\Gamma$, $T_c$	198–1,000+	(Wang et al., 2022)
Interfaces/Nanocrystals	$\gamma_{hkl}$, $E_b$, $ΔE_f$	10–100s	(Mathew et al., 2016)
Organic OPV	DFT/ML PCE, HOMO/LUMO, ML cal. uncertainty	91,000–2.3e6	(John et al., 2018, Pyzer-Knapp et al., 2015)

For example, the DBCata deep generative model achieves $<$0.1 eV adsorption-energy fidelity for >90% of 15,911 O/OH/metal-alloy slab structures in minutes of GPU time, compared to weeks for DFT-only optimization (Huo et al., 17 Dec 2025). In thermoelectrics, descriptors like $t$ and $\chi$ reproducibly retrieve known $zT$ champions and predict new materials, validated against experiment (Xing et al., 2017, Jia et al., 2019). High-throughput structure–property maps for MOFs in gas separation result in experimentally validated top membranes (Afzal et al., 2019).

5. Model Calibration, Uncertainty Quantification, and Feedback

As HTES predictions progress toward experiment, model calibration and validation become crucial:

• Bayesian Calibration: Gaussian-process calibration over extended-connectivity fingerprints effectively removes systematic DFT or model biases and yields quantitative uncertainties per structure. This empowers uncertainty-aware selection and resource allocation in HTES pipelines (Pyzer-Knapp et al., 2015).
• Active Feedback Loops: Experimental data are cycled back (e.g., via weighted cross-validation or retraining), both updating the model and flagging regions where predictions are extrapolative or uncertain.
• Automated Outlier Detection: Hybrid chemical heuristics and GNN-based anomaly classification ensure >94% adsorption-energy fidelity in direct DFT screening (Huo et al., 17 Dec 2025); error networks estimate prediction reliability in device or property regression tasks (Hußner et al., 2023).
• Comparison to Known Data: Cross-sectional validation against established leaders (e.g., GeTe, PbTe for TE; PCE for organic solar; Li₁₀GeP₂S₁₂ for Li-ion conduction) is standard (Xing et al., 2017, Kahle et al., 2019).

6. Emerging Capabilities, Limitations, and Generalization

HTES development is advancing on several fronts:

• ML/Physics Hybridization: Incorporating equivariance, chemical priors, and uncertainty in generative frameworks enables robust structure optimization beyond the training set (e.g., DBCata's Brownian-bridge/PaiNN architecture) (Huo et al., 17 Dec 2025).
• Extension to Complex and Multi-objective Materials Spaces: Ionic substitution rules, surrogate interpolation, and evolutionary algorithms scale composition space to tens of thousands of candidates in PV and photocatalytic domains (Stafford et al., 2023, Liu et al., 2022).
• Efficient Phase Stability Assessment: Energy hull analysis and rapid convex-hull construction in tandem with accurate surrogate energetics ensure only experimentally relevant candidates progress (Armiento et al., 2013, Wang et al., 2022).
• Limitations: Single-band or rigid-band approximations may miss multiband or nonparabolic behavior; some surrogate descriptors (e.g., $\lambda_\Gamma$) may fail outside their validated structural class; domain shift in ML models requires retraining on sufficiently diverse or updated simulated data (Deng et al., 2021, Xing et al., 2017, Hußner et al., 2023, Huo et al., 17 Dec 2025).
• Experimental Integration: HTES-guided synthesis, rapid device characterization via ML-extracted transport parameters, and “resurrection” of previously discarded compositions from legacy datasets showcase the increasing relevance of loop-closed pipelines (Hußner et al., 2023, Afzal et al., 2019).

7. Impact and Pathways to Rational Materials Discovery

HTES has demonstrated a transformative impact on the pace and efficiency of materials innovation. For catalysis, rapid, descriptor-based screening and subsequent experimental validation (e.g., Pd–Fe, Pd–Zn for PEMFCs, Cu-doped MoS₂ for ORR) have led to commercial-grade discoveries (Afzal et al., 2019). In energy storage and conversion, the deployment of robust surrogate and calibration methods systematically reduces the required DFT budget by orders of magnitude, unlocks exploration of 10^6–10^9 chemical permutations, and aligns computational predictions with real-world device performance.

A plausible implication is that continued convergence of physics-driven surrogate models, scalable data infrastructure, and real-time experiment–theory feedback will further compress the time-to-discovery in energy materials, while raising the reliability and quantitative predictive power of computational pipelines (Afzal et al., 2019, Huo et al., 17 Dec 2025, Hußner et al., 2023).