PaperlayerView exact revision on arXiv

Source-linked AI summary

AdsorbML: A Leap in Efficiency for Adsorption Energy Calculations using Generalizable Machine Learning Potentials

Janice Lan, Aini Palizhati, Muhammed Shuaibi, Brandon M. Wood, Brook Wander, Abhishek Das, Matt Uyttendaele, C. Lawrence Zitnick, Zachary W. Ulissi

arXiv:2211.16486v3cond-mat.mtrl-scics.LG

TL;DR

Identifying low-energy adsorbate-surface configurations is difficult because many sites and orientations must be considered, and intuition does not scale to many systems. The paper uses ML-potential relaxations followed by DFT refinement, reporting that current ML models accelerate DFT and enable more accurate estimates for global-optimization properties.

  • Problem

    Finding the lowest-energy adsorbate-surface configuration is difficult amid many sites and orientations, while intuition and trial-and-error do not scale to many systems.

    Source: INTRODUCTION · pp. 1–2

  • Method

    The paper samples heuristic and random configurations, relaxes them with ML potentials, then refines the best-k energies with single-point or full-relaxation DFT.

    Source: INTRODUCTION · p. 3

  • Results

    Current state-of-the-art ML models accelerate DFT calculations for catalysis and enable more accurate estimates of properties requiring global optimization, including adsorption energies.

    Source: INTRODUCTION · p. 1, DISCUSSION · p. 9

  • Takeaways & Limitations

    Fine-tuning strategies can help enable applications beyond idealized adsorbate-surface catalysts, including to metal-organic frameworks and zeolites.

    Source: DISCUSSION · p. 9

    ML relaxations without subsequent DFT are still not accurate enough for practical applications.

    Source: DISCUSSION · p. 9

Abstract

from arXiv · show

Computational catalysis is playing an increasingly significant role in the design of catalysts across a wide range of applications. A common task for many computational methods is the need to accurately compute the adsorption energy for an adsorbate and a catalyst surface of interest. Traditionally, the identification of low energy adsorbate-surface configurations relies on heuristic methods and researcher intuition. As the desire to perform high-throughput screening increases, it becomes challenging to use heuristics and intuition alone. In this paper, we demonstrate machine learning potentials can be leveraged to identify low energy adsorbate-surface configurations more accurately and efficiently. Our algorithm provides a spectrum of trade-offs between accuracy and efficiency, with one balanced option finding the lowest energy configuration 87.36% of the time, while achieving a 2000x speedup in computation. To standardize benchmarking, we introduce the Open Catalyst Dense dataset containing nearly 1,000 diverse surfaces and 100,000 unique configurations.

INTRODUCTION

Adsorption energies are key inputs to catalyst screening, but finding the global minimum across many adsorbate placements is difficult and costly with DFT. The paper proposes combining machine-learning potentials with DFT to search configurations more efficiently, while introducing a dense benchmark for evaluating such methods.

INTRODUCTION

  • Adsorption energies provide starting points for free-energy diagrams used to identify favorable catalyst reaction pathways.Adsorption energies of reaction intermediates can also serve as descriptors correlated with experimental activity or selectivity.Source: INTRODUCTION · p. 1
  • Finding the global minimum requires exploring multiple binding sites and molecular orientations, while DFT relaxation reaches only a local minimum.DFT iteratively uses per-atom forces to update positions until a local energy minimum is reached.Source: INTRODUCTION · p. 1
  • Valid adsorption energies require relaxed structures without desorption, adsorbate dissociation, or significant surface mismatch.Such outcomes can occur without indicating an inaccurate relaxation, but they yield invalid or inaccurate adsorption energies as defined here.Source: INTRODUCTION · p. 2
  • Expert intuition and common surface-symmetry heuristics do not scale well as surface and adsorbate complexity grows, especially for flexible adsorbates.Flexible adsorbates have multiple configurations, increasing the system’s effective degrees of freedom.Source: INTRODUCTION · p. 2
  • A single full DFT relaxation can take about 24 hours, and exploring many configurations can extend the computation to days or weeks.The paper estimates that brute-force CO2RR screening of 100,000 materials under its stated assumptions would require about 4.5 billion CPU-days.Source: INTRODUCTION · p. 2
  • OC20-Dense addresses the lack of standardized validation data by densely sampling configurations for held-out adsorbate-surface combinations across four OC20 subsplits.OC20’s single relaxed configuration per combination is unlikely to be the global minimum, making it unsuitable for finding minimum binding energies.Source: INTRODUCTION · p. 3
  • The proposed hybrid approach relaxes heuristic and random configurations with ML potentials, then refines the best-k energies using DFT single-points or full relaxations.The two refinement options provide different accuracy-efficiency trade-offs.Source: INTRODUCTION · p. 3

Related Work

Prior searches have used DFT-based global optimization, active learning, or bespoke models, while AdsorbML applies generalizable ML potentials to adsorption-energy search. The section also introduces OC20-Dense as a benchmark and evaluates previously published GNN models on it.

Related Work

  • Prior approaches include constrained minima hopping with DFT and active learning with a Gaussian process, while earlier ML accelerators often used bespoke models for each adsorbate-catalyst pair.The paper presents generalizable ML potentials as a way to broaden applicability while reducing human and computational cost.Source: Related Work · p. 3
  • AdsorbML targets adsorption-energy search across a spectrum of accuracy-efficiency trade-offs.The work also benchmarks literature GNNs on OC20-Dense and identifies promising models for practical screening.Source: Related Work · p. 3

OC20-Dense Evaluation

  • OC20-Dense benchmarks adsorption-energy search; each split comprises ∼1,000 unique adsorbate-surface combinations spanning 74 adsorbates and 800+ inorganic bulk crystal structures, with 80,000+ heuristically and randomly generated configurations.The dataset required ∼4 million CPU-hrs of compute to complete.Source: OC20-Dense Evaluation · p. 3, Related Work · p. 3
  • The evaluation covers previously published SchNet, DimeNet++, PaiNN, GemNet, and SCN model variants; the models were not trained for this work.The evaluated set includes models trained on OC20 and its accompanying ab initio molecular-dynamics dataset.Source: OC20-Dense Evaluation · p. 3
  • Each OC20-Dense validation and test split samples about 250 systems from each of ID, OOD-Adsorbate, OOD-Catalyst, and OOD-Both.The split sizes are summarized in Table II.Source: Table II · p. 4
  • The benchmark’s DFT-Heuristic+Random baseline adds uniformly sampled surface sites and random adsorbate orientations to heuristic configurations for a more thorough search.The paper describes this as a more accurate adsorption-energy estimate than the heuristic-only baseline, though computationally more expensive.Source: OC20-Dense Evaluation · p. 4

ML Relaxations

ML relaxations are evaluated against OC20-Dense using a tolerance-based success criterion and DFT checks of predicted structures. The reported top-model success rates motivate hybrid ML+DFT strategies, which trade additional compute for improved success rates.

ML Relaxations

  • ML relaxations start from identical heuristic and random configurations, and structures violating adsorption constraints are removed from evaluation.The excluded cases include dissociation, desorption, and surface mismatch.Source: ML Relaxations · p. 4
  • 56.52% and 48.03% success rates were achieved by eSCN-MD-Large and GemNet-OC-MD-Large, respectively, on OC20-Dense.Success requires an energy within 0.1 eV of or lower than the DFT adsorption energy, plus a DFT single-point check within 0.1 eV of the ML-predicted energy.Source: ML Relaxations · p. 4
  • Energy MAE between ML and DFT adsorption energies is reported alongside success rates and correlates with them and OC20 S2EF metrics.The table also reports OC20 S2EF Val-ID results.Source: ML Relaxations · p. 5
  • Hybrid ML+DFT strategies are explored to improve success rates at the cost of additional compute.The motivation is that a success rate near 50% could leave high-throughput screening candidates requiring further analysis as false positives.Source: ML Relaxations · p. 5

AdsorbML Algorithm

AdsorbML ranks ML-relaxed configurations and verifies the best k with DFT, offering single-point and full-relaxation strategies. The experiments report substantial accuracy–efficiency trade-offs, with results varying by strategy, model, and configuration source.

AdsorbML Algorithm

  • AdsorbML ranks ML-relaxed configurations and sends the lowest-energy k to DFT for single-point evaluation or full relaxation, then selects the minimum DFT energy.ML energies rank configurations, while DFT supplies the final energy; invalid structures are filtered during relaxation.Source: AdsorbML Algorithm · p. 5, Figure 2 · p. 6

Experiments

  • At k = 3, eSCN-MD-Large ML+SP reaches 87.36% success and 2296x speedup; Figure 3 maps the accuracy-efficiency trade-offs across methods.GemNet-OC-MD-Large reaches 84.43% success and 2299x speedup at the same k.Source: Experiments · p. 5, Figure 3 · p. 6
  • At k = 5, GemNet-OC-MD-Large ML+RX reaches 91.61% success and eSCN-MD-Large reaches 90.60%; at k = 3, their speedups are 288x and 356x, respectively.The corresponding k = 5 speedups are 172x for GemNet-OC-MD-Large and 215x for eSCN-MD-Large.Source: Experiments · pp. 5–7
  • For the best models, ML+SP is almost 8x faster than ML+RX with only a 3–4% decrease in success rate.The paper reports that this difference is much larger for weaker models.Source: Experiments · p. 7
  • At k = 5 ML+RX, eSCN-MD-Large finds much lower minima for 9.10% of systems, versus 15.57% for DimeNet++; the most accurate models do not necessarily find more such minima.The authors suggest that some model noise may aid in finding better minima.Source: Experiments · p. 7
  • Across OC20-Dense splits, eSCN-MD-Large ML+SP at k = 5 has 86.00% success on ID and 88.35% on OOD-Both; the authors attribute the discrepancy to 250 systems per split.Figure 4 shows success rates across dataset splits and models.Source: Figure 4 · p. 7, Experiments · p. 8
  • At k = 5, SCN-MD-Large ML+SP success falls from 87.77% with both configuration sources to 82.94% with random-only and 62.18% with heuristic-only configurations.The authors report that random configurations can have a larger impact.Source: Experiments · p. 8

DISCUSSION

AdsorbML offers compute-dependent ML+DFT trade-offs for screening, while several practical boundaries remain: enumeration is not exhaustive, and ML-only relaxations are not yet accurate enough for practical use. The discussion also identifies better minima from ML relaxations and evaluation constraints as areas to consider.

DISCUSSION

  • ML+SP at k = 2 achieves an 85% success rate while screening ∼3400x more materials than DFT alone under a fixed compute budget.For deeper study, ML+RX fully optimizes structures with DFT while reducing the required relaxation steps.Source: DISCUSSION · pp. 8–9
  • ML models found better minima 5%-15% of the time, depending on efficiency trade-offs, a pattern the authors link to force-prediction noise and identify for further study.They propose exploring how implicit and explicit noise affect ML relaxations and downstream success rates.Source: DISCUSSION · p. 9
  • The enumerated initial configurations span more than traditional heuristics but are not exhaustive, and increasing random DFT configurations from 80% to 100% changed success rate by only 1.6%.Alternative global-optimization and initial-configuration methods, as well as duplicate-structure handling, are suggested for further work.Source: DISCUSSION · p. 9
  • ML relaxations without subsequent DFT are not accurate enough for practical applications, and OC20 validation data should not train models evaluated on OC20-Dense because the validation sets overlap.The authors also encourage DFT single-point evaluation on OC20-Dense validation, since predicting only low energies could manipulate success rate.Source: DISCUSSION · p. 9
  • The models are best suited to idealized adsorbate-surface catalysts and were trained with revised PBE without spin-polarization; generalization to other chemistries or theory levels may require fine-tuning or other training strategies.Generalizing to more diverse materials and molecules without relying on DFT remains a community challenge.Source: DISCUSSION · p. 9
  • The authors conclude that current state-of-the-art ML models can accelerate DFT calculations for catalysis and improve estimates of properties requiring global optimization, such as adsorption energies.They frame the work as an initial step toward reducing DFT cost in catalysis and computational chemistry more broadly.Source: DISCUSSION · p. 9

METHODS

OC20-Dense extends OC20 to evaluate adsorption-energy estimation using densely sampled adsorbate configurations. Its validation and test splits cover diverse subsplits and combine heuristic and random initial-configuration strategies.

METHODS

  • OC20-Dense addresses the limitation that OC20 often samples only one randomly selected adsorption configuration per adsorbate-surface combination, which may not represent that combination's minimum energy.The dataset is designed to investigate models' ability to find adsorption energies across densely sampled initial configurations.Source: METHODS · p. 9
  • For each split, the dataset samples about 1,000 adsorbate-surface combinations and draws uniformly from ID, OOD-Adsorbate, OOD-Catalyst, and OOD-Both subsplits.Heuristic configurations enumerate symmetry-equivalent sites with adsorbate rotations and wobbles; the random strategy samples M=100 surface sites, subject to filtering.Source: METHODS · p. 10
  • The OC20-Dense test set contains 989 combinations spanning 74 adsorbates and 837 bulks, with 56,282 heuristic and 49,432 random configurations relaxed using DFT.After filtering, the dataset averages 56 heuristic and 50 random configurations per system and required ∼4 million compute hours.Source: METHODS · p. 10

Evaluation Metrics

The evaluation uses success rate and DFT speedup against specified baselines, while filtering relaxed structures that do not represent valid adsorption energies. The supplied result tables identify evaluation sets and baselines but provide no cell values here.

Evaluation Metrics

  • Success rate counts systems whose strategy returns an energy within σ, or lower, of the DFT adsorption energy; the selected tolerance is σ = 0.1 eV, and DFT speedup compares electronic-step counts against DFT-Heur+Rand.Electronic steps were chosen because they correlated better with compute time than ionic relaxation steps; hardware-dependent compute time and ML relaxation costs are excluded.Source: Evaluation Metrics · p. 10

Relaxation Constraints

  • The analysis discards desorption, dissociation, and significant adsorbate-induced surface changes because these outcomes do not represent the intended adsorption energy or can invalidate clean-surface referencing.Detection methods use neighborhood connectivity, distances, and covalent radii; thresholds may be tightened for particular applications.Source: Relaxation Constraints · pp. 10–11

Main Paper Results

  • Supplementary Tables I–IV cover success and speedup results across k for OC20-Dense test and validation sets, evaluated against DFT-Heuristic+Random or DFT-Heuristic baselines.Tables I–II use the more exhaustive DFT-Heuristic+Random baseline; Tables III–IV use the more common DFT-Heuristic baseline.Source: Table I · p. 14, Table II · p. 14, Main Paper Results · p. 14

Subsplit Results

The subsplit results are organized by OC20-Dense validation and test sets and by the ML+SP and ML+RX strategies. The supplied table descriptions specify the comparisons but not their outcomes.

Subsplit Results

OOD-Both

  • The OOD-Both table descriptions organize success-rate results by subsplit, ML+SP or ML+RX strategy, and k for the OC20-Dense test and validation sets.The captions identify the DFT-Heur+Rand comparison but do not state which strategy or k performs best.Source: Table V · p. 16, Table VI · p. 17, OOD-Both · p. 17

Model Implementation and Compute Details

The section reports how ML inference affects total speedup and evaluates whether deduplicating relaxed configurations improves best-k selection. It also documents the models and compute settings used.

Model Implementation and Compute Details

  • The study used seven pretrained GNN models without training new models, with identical optimization settings and up to 300 relaxation steps per model.Relaxations stopped earlier if the maximum per-atom force norm reached 0.02 eV/Å; GemNet-OC-MD-Large used only the force variant.Source: Model Implementation and Compute Details · p. 18
  • 3596x and 3885x ML+SP speedups at k = 1 fall to 1147x and 1686x when ML compute is included for SCN-MD-Large and GemNet-OC-MD-Large, respectively.For ML+RX, the largest models see only 24.6% and 14.2% slowdowns; ML compute matters less as k increases to 5.Source: Model Implementation and Compute Details · p. 18
  • The compute-time comparison uses an alternative speedup based on total runtime, reporting results with and without ML runtime for ML+SP and ML+RX at k = 1 and k = 5.ML relaxations were run on 32GB NVIDIA V100 cards; the reported speedup metrics otherwise count DFT electronic steps only.Source: Model Implementation and Compute Details · p. 18, Table VIII · p. 19

Deduplication

  • Deduplication clusters ML-relaxed configurations by energy difference, then selects the lowest-energy candidate from each cluster until k placements are chosen.The threshold ∆E controls clustering: too small a value may deduplicate little, while too large a value may group distinct systems.Source: Deduplication · p. 19, Table IX · p. 19
  • Deduplication produced only marginal gains across k; ∆E = 0.01 eV gave a minor improvement over no deduplication.The authors identify alternative strategies, such as structure-based deduplication, or increasing the number of placements as future directions.Source: Deduplication · p. 19

Varying Heuristic+Random ratios

The section examines how random placements, relaxation choices, and ML ranking affect AdsorbML results on OC20-Dense validation. It reports diminishing returns from adding random configurations and compares alternative configuration and ranking baselines.

Varying Heuristic+Random ratios

  • 1.6% separates the success rates at 80% and 100% random configurations, compared with an 8% improvement from 0% to 10% additional random configurations.Success rates increase as the proportion of random configurations grows, but gains diminish at higher proportions.Source: Varying Heuristic+Random ratios · p. 20, Table X · p. 20

Additional Results

  • Relaxations and single-points have similar success rates, but the SCN-MD-Large energy-difference histogram shows lower predicted energies for relaxations.The figure reports validation-set results at k = 5 and marks differences within 0.1 eV as comparable successes.Source: Additional Results · p. 20, Figure 1 · p. 21

Configuration analysis

  • Random-only configurations perform slightly worse and heuristic-only configurations significantly worse against the combined ground truth; heuristic-only becomes competitive against heuristic-only ground truth.The table comparison indicates that random configurations help both AdsorbML and ground truth.Source: Configuration analysis · p. 20, Table XI · p. 22

Random baselines

  • Randomly selected k placements perform badly, while selecting the k placements with the highest predicted energies performs worst; the regular best-k ranking is the reference.These comparisons use SCN-MD-Large single-points on the OC20-Dense validation set.Source: Table XII · p. 22, Random baselines · p. 22

SUPPLEMENTARY FIGURES

The supplementary material shows an accuracy-speedup comparison and describes physical filters for relaxed structures. It also reports how often ML and DFT relaxations violate those constraints.

OC20-Dense Validation Success v. Speedup

  • 86.33% success and 1331x speedup mark the balanced option in the validation-set accuracy-efficiency comparison.The figure plots success rate against DFT speedup for ML+RX and ML+SP across best-k values and baseline GNN models.Source: Figure 2 · p. 23

Relaxation Constraints

  • Dissociation, desorption, and surface mismatch are filtered using neighborhood-based checks of connectivity before and after relaxation.For desorption detection, a 1.5 multiplier on covalent radii is used so that systems are discarded only when the adsorbate has no surface interaction, avoiding removal of physisorbed systems.Source: Relaxation Constraints · pp. 23–24

Constraint Counts

  • Top-performing models such as SCN-MD-Large and GemNet-OC have fewer constraint-violating relaxations removed, with counts broken down by violation type for ML and DFT relaxations.A system may have more than one violation, so the total need not equal the sum of violation-type counts.Source: Constraint Counts · p. 24, Table XIII · p. 24

DFT and Calculation Details

The section specifies the DFT calculation setup and gas-phase reference convention, then describes differences in OC20-Dense validation and test placement procedures. The changelog records dataset and evaluation updates across document versions.

DFT and Calculation Details

  • DFT calculations used VASP with PAW pseudopotentials and the RPBE functional, with limits of 60 electronic steps for relaxations and 300 for single-points.Unconverged single-points and unconverged electronic steps at relaxed structures were discarded.Source: DFT and Calculation Details · p. 24
  • In this work, Egas is often a linear combination of gas-phase reference species rather than the stable gas-phase adsorbate itself.The passage notes that referencing a stable gas-phase molecule may be less helpful for thermodynamically consistent free-energy diagrams.Source: DFT and Calculation Details · p. 24

OC20-Dense Placement Details

  • OC20-Dense validation and test sets were generated at different stages with different placement code; the passage reports that model trends were retained across splits.The improved test-set heuristic strategy is reflected in DFT-Heur baselines of 87.76% success and 1.81x speedup, versus 71.12% and 2.87x for validation.Source: OC20-Dense Placement Details · pp. 24–25

CHANGELOG

  • The changelog records updated compute times, speedups, dataset statistics, a new OC20-Dense test set, and evaluation of eSCN-MD-Large.Version 3 records publication in npj Comput. Mater.Source: CHANGELOG · p. 26

Related papers

Loading 2211.16486v3…