Pith. sign in

REVIEW 4 major objections 4 minor 44 references

HPC-AI Coupling Methodology for Scientific Applications

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proposes that HPC-AI integration reduces to three reusable coupling patterns—surrogate, directive, and coordinate—and demonstrates them in materials science with orders-of-magnitude speedups.

desk verdict A useful taxonomy of HPC-AI coupling patterns, but the coordinate-pattern case study has a load-bearing confidence-calibration and counting problem that undercuts the main performance claim. read the letter →

arxiv 2507.01025 v1 pith:LLKEI37M submitted 2025-06-17 cs.CE cs.AIphysics.comp-ph

Yutong Lu , Dan Huang , Pin Chen This is my paper
classification cs.CEcs.AIphysics.comp-ph
keywords highperformancecomputingartificialintelligencesurrogatepatterndirectivecoordinatematerialssciencedensityfunctionaltheorycrystalstructureprediction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Scientists increasingly combine high-performance computing (HPC) and AI, but each integration is built from scratch. This paper argues that the combinations can be organized into exactly three patterns—surrogate, directive, and coordinate—and that naming the pattern is the first design decision for any HPC-AI ensemble. In the surrogate pattern an AI model replaces part of a physics simulation; in the directive pattern AI steers the simulation's search; in the coordinate pattern AI and simulation exchange feedback iteratively, with an agent deciding which jobs go where. The paper builds one concrete system per pattern in materials science and reports large speedups, including an AI surrogate that is roughly 886 times faster than PBE-level DFT and about 8,245 times faster than HSE-level DFT at comparable accuracy, and a coordinate-pattern agent that produced 14 Fe2O3 candidate structures in 50 iterations using only five DFT runs. If the taxonomy holds, other scientific fields can use it as a blueprint for designing their own HPC-AI systems.

What carries the argument

The load-bearing mechanism is the three-pattern taxonomy itself, each pattern formalized by an objective function: surrogate minimizes the discrepancy between $f_{\mathrm{AI}}$ and $f_{\mathrm{HPC}}$; directive solves for the optimal input $x^* = \arg\min_x L(f_{\mathrm{HPC}}(x), y_{\mathrm{target}})$; coordinate iterates $z_{t+1}=f_{\mathrm{AI}}(x_t)$, $y_{t+1}=f_{\mathrm{HPC}}(z_{t+1})$ to minimize $L(f_{\mathrm{HPC}}(f_{\mathrm{AI}}(x)), y_{\mathrm{target}})$. Within the implementations, the paper relies on two named models built for the demonstrations: Crysformer, an SE(3)-equivariant graph attention network that maps crystal structures to properties using spherical-harmonic edge embeddings and equivariant gates, and EquiCSP, a joint diffusion model over lattice, fractional coordinates, and composition. The coordinate pattern adds a decision rule: an ensemble or Monte-Carlo-dropout variance $\hat{\sigma}^2$ is compared to a threshold $\tau$, and a Match Discriminator probability $p_{\mathrm{match}}$ scores generated structures, so only low-confidence cases go to DFT. That confidence gating is what converts the abstract coordinate loop into a concrete compute-routing mechanism.

What would settle it

Take a random sample of structures with DFT-computed formation energies, run the Crysformer ensemble to get predicted means and variances, and plot the actual error against the declared 90% confidence region. If the accepted-by-threshold predictions have an error above the 0.12 eV/atom bound substantially more (or less) often than 10%, the threshold is miscalibrated and the coordinate pattern's compute savings would be either unsafe or overly conservative.

Watch

Extended reading notes

Core claim

The paper's central claim is that the ad hoc practice of coupling HPC simulations with AI can be systematized into three distinct patterns, each with a formal objective. In the surrogate pattern, an AI model $f_{\mathrm{AI}}$ is trained to approximate the high-fidelity simulation $f_{\mathrm{HPC}}$ by minimizing $\mathbb{E}_{x\sim p(x)}\lVert f_{\mathrm{HPC}}(x)-f_{\mathrm{AI}}(x)\rVert^2$, replacing the expensive simulation outright. In the directive pattern, AI acts as an optimizer that solves $x^* = \arg\min_x L(f_{\mathrm{HPC}}(x), y_{\mathrm{target}})$, guiding the simulation toward promising parameter regions. In the coordinate pattern, the two interact through an iterative loop $z_{t+1}=f_{\mathrm{AI}}(x_t)$, $y_{t+1}=f_{\mathrm{HPC}}(z_{t+1})$ with the objective $\min L(f_{\mathrm{HPC}}(f_{\mathrm{AI}}(x)), y_{\mathrm{target}})$, and a confidence mechanism routes jobs between AI and DFT. The materials-science demonstrations are offered as evidence that these patterns are not just abstract categories: Crysformer achieves state-of-the-art formation-energy and bandgap errors, EquiCSP generates about 5,663 candidate structures per hour with about 1,859 validated, and the coordinate-pattern LLM agent discovers 14 Fe2O3 structures below 1 eV/atom while avoiding nine DFT calculations. The conclusion the author draws is that these patterns, though exemplified in materials science, provide generally applicable guidance for future HPC-AI ensembles.

Load-bearing premise

The coordinate pattern's efficiency claim rests on the assumption that an AI prediction with estimated confidence above the 90% threshold is actually accurate enough to be trusted without a DFT calculation; the threshold is set from published error ranges rather than from a calibration experiment on the exact models and data used here.

Editorial extensions

If this is right

  • Surrogate-pattern models can replace DFT for rapid screening at two to four orders of magnitude lower cost, with accuracy close to the training reference (0.020 eV/atom formation-energy MAE on the benchmark set).
  • Directive-pattern generative models can produce millions of candidate crystal structures, but the hybrid sampling strategy—learned prior first, random sampling later—matters: valid-structure rate was 43% initially dropping to 22% for prior sampling, versus a stable ~28% for random sampling.
  • Coordinate-pattern workflows can cut DFT calls substantially: in the 50-iteration Fe2O3 run, 9 of 14 accepted results came from AI predictions and only 5 required DFT, yielding 14 valid low-energy structures.
  • The three-pattern comparison table (generalization, efficiency, applicability) gives designers a decision rule: use surrogate when speed dominates and data coverage is adequate, directive for optimization-centric searches, and coordinate for iterative high-stakes workflows.
  • Because the patterns are defined by data flow rather than domain, they can be transferred to any field pairing numerical simulation with learned models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The confidence threshold $\tau$ in the coordinate pattern is a tunable policy parameter; one testable extension is to learn $\tau$ online from the observed DFT errors on low-confidence calls, which would make the compute savings self-calibrating rather than fixed at 90%.
  • The surrogate's reported accuracy is measured on in-distribution test sets; an editorial implication is that on out-of-distribution or rare structures the speedup could erode unless the confidence gating of the coordinate pattern is attached.
  • The taxonomy is one of three possible axes; a neighboring design question is how to schedule retraining or fine-tuning of the AI model within each pattern, which the paper leaves mostly static for surrogate and directive patterns.
  • The Match Discriminator for structure confidence is a learned classifier on RMSD-based labels; an editorial test is whether its probability is calibrated enough to serve as a reliable gate when the generator is asked to extrapolate to new chemistries.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. This paper proposes a methodology for coupling high-performance computing (HPC) and artificial intelligence (AI) in scientific applications, organized around three patterns: surrogate, directive, and coordinate. In the surrogate pattern, AI models replace DFT calculations; in the directive pattern, a diffusion generative model proposes crystal structures that are then validated by DFT; in the coordinate pattern, an LLM-based agent decides when to trust AI property predictions and when to invoke DFT, with a confidence threshold controlling the tradeoff. The manuscript presents mathematical notations for each pattern, reports a Crysformer model for property prediction, an EquiCSP-based structure generation workflow, and a Fe2O3 design case study, and claims significant performance improvements and broad applicability beyond materials science.

Significance. The three-pattern taxonomy is a useful organizing framework for the growing body of HPC-AI systems, and the paper gives concrete, named implementations of each pattern. The strength of the paper is conceptual: the definitions in Section 2 are clear, the figures illustrate the patterns well, and the case studies demonstrate feasibility of all three coupling modes. If the performance claims were properly validated, the work would be a valuable reference for designers of HPC-AI ensembles. As presented, however, the quantitative support is thin: the state-of-the-art claim in Table 1 lacks statistical error bars, the directive-pattern results have no baseline comparison, and the coordinate-pattern efficiency claim rests on an uncalibrated confidence threshold and internally inconsistent counts. The taxonomy itself is not invalidated by these issues, but the empirical evidence does not yet establish the claimed 'significant performance improvements'.

major comments (4)
  1. [§3.3, §4.3] The coordinate-pattern efficiency claim rests on a confidence threshold that is neither calibrated on the models used nor consistently defined. In Section 3.3 (steps 3–5) the decision criterion is a variance threshold (σ^2 < τ) or match-probability threshold (p_match > τ), while Section 4.3 reports a '90% confidence' threshold and justifies it by citing global formation-energy MAEs of 0.081–0.136 eV/atom from Xie and Grossman (2018). A global MAE does not determine a per-prediction 90% confidence region, no mapping is given between the variance threshold and the 90% probability, and no reliability diagram or calibration experiment on Crysformer's actual confidence scores for the generated Fe2O3 structures is provided. The manuscript itself notes in Section 5.1 that surrogate patterns need confidence evaluation for reliability, but the coordinate case study does not supply that evaluation. If the confidence scores are overconfident, the workflow returns unreliable formation energies; if underconfident, it saves little over all-DFT. Because this is the only end-to-end quantitative demonstration of the methodology, the central performance claim is not established by the reported experiments.
  2. [§4.3, Figure 11] The reported counts do not reconcile. Section 4.3 states that the AI model was applied nine times, that five DFT calculations were performed for low-confidence predictions, and that the workflow 'effectively circumvented nine additional time-consuming DFT computations.' If each AI application replaces a DFT run, the number of circumvented DFT calls is at most nine, and the net saving relative to an all-DFT baseline is four, not nine; the sentence could also be read as claiming nine DFT calls were avoided in addition to the five performed, which would require 14 AI applications. The baseline is never defined—that is, how many of the 50 iterations would require DFT in the absence of AI is not stated—so no valid efficiency gain can be quantified from the reported counts.
  3. [§4.1, Table 1, Figure 5] The 'state-of-the-art' claim in Section 4.1 is stronger than the evidence supports. Crysformer's formation-energy MAE (0.020 eV/atom) differs from Matformer (0.021) by 0.001 eV/atom and from ALIGNN (0.022) by 0.002 eV/atom, and the bandgap MAE differs from Matformer by only 0.004 eV; no error bars, number of seeds, or run-to-run variance are reported. The training-size curve in Figure 5 also has no error bars. The qualitative conclusion that the surrogate pattern is viable is not in doubt, but the quantitative ranking should be reported with uncertainty estimates or the wording should be softened.
  4. [§3.2, §4.2, §5.2] The directive-pattern case study does not establish the efficiency and 'reduced unproductive evaluations' attributed to it in Section 5.2. The reported throughput (5,663 structures/hour, 1,859 valid/hour) and the validity-rate curves in Figure 6 describe the generative model, but there is no comparison with a non-AI baseline such as random search or evolutionary CSP, and no accounting for the computational cost of the high-throughput DFT validation step. Without such a baseline, the claim that AI guidance focuses computational effort on high-value regions is plausible but unsupported by the experiments.
minor comments (4)
  1. [§3.1, Eq. (9)] The notation T_c(i) in Eq. (9) is not defined; it appears to be a property-specific readout at node i, but the reader cannot determine this from the text.
  2. [§3.2, Eq. (13)] The loss weights λ_L, λ_F, and λ_A are introduced, but values are given only for the CSP task (λ_A = 0); the 'larger value' used for the ab initio generation task is never specified, so the training objective is not reproducible.
  3. [Figure 5] The text says training sizes were selected at intervals of 5,000, but the plotted curve appears to show only four points; the exact sample sizes and the number of runs per point should be stated.
  4. [Table 2] Table 2 compares Crysformer against DFT values from MP, Matgen, AFLOW, and OQMD, but the test-set composition and the criteria for matching structures across these databases are not described, which makes the comparison difficult to interpret.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pattern definitions are definitional, the case-study claims are empirical, and the confidence-threshold concern is a calibration gap rather than a self-referential reduction.

full rationale

The paper does not derive its central claims from quantities that it also fits. Section 2 defines surrogate, directive, and coordinate patterns with standard composition and optimization equations; these are formal definitions of a taxonomy, not derivations of a predicted constant from a fitted parameter. The surrogate-pattern results are benchmarked on MP test data and compared with independent models, and the time ratios (886x, 8245x) are measured runtimes rather than identities. The directive-pattern numbers come from an existing generative framework, EquiCSP, which is self-cited but is used as a component; the validity of the proposed taxonomy does not reduce to EquiCSP's validity. The coordinate pattern's 90% confidence threshold is justified by reported external formation-energy MAEs rather than by fitting the threshold to the observed DFT outcomes; while this is an unvalidated calibration assumption and a legitimate correctness risk, it is not a case where the claimed savings are true by construction. Likewise, the internal inconsistency in the number of circumvented DFT computations (nine AI predictions and five DFT calls reported as 'circumvented nine' DFT runs) is an accounting and empirical-support problem, not circular reasoning. No quoted equation or self-citation chain forces the paper's conclusions to equal its inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities, particles, forces, or dimensions are introduced. The proposed Match Discriminator is a machine-learned component, not a new scientific entity, and it has no independent falsifiable handle outside this workflow. The free parameters are the confidence threshold tau, the undefined RMSD threshold d, and the diffusion loss weights.

free parameters (3)
  • Confidence threshold tau = 0.90 (90%)
    Chosen in Section 4.3 based on reported DFT formation energy MAEs from Xie and Grossman (2018) and Chen et al. (2019), not from a calibration experiment on the models and data used in this paper.
  • RMSD match threshold d = not specified
    Used to label generated structures as matches for training the Match Discriminator in Section 3.3; the value is left undefined.
  • Diffusion loss weights = lambda_L = 1, lambda_F = 1, lambda_A = 0 for CSP; larger lambda_A for ab initio generation
    Weights in Equation (13) are set to balance loss scales; the specific larger value for lambda_A is not given.
assumptions (4)
  • domain assumption DFT calculations are the ground truth for training surrogate models and for validating generated structures.
    The paper assumes first-principles DFT is the reference accuracy, which is standard in materials informatics but not proven for all properties.
  • domain assumption Validity filters (component count, electroneutrality, symmetry, similarity) define potentially valid crystal structures.
    Section 4.2 lists four filters used to accept generated structures; these are heuristic and can bias the reported valid-generation rate.
  • domain assumption Low predictive variance from MC dropout or ensembles indicates high confidence.
    Section 3.3 uses predictive variance as a confidence score; this is a standard uncertainty-quantification assumption but is not validated here.
  • standard math Standard function composition and L2 norms are appropriate for defining coupling patterns.
    The mathematical framework in Section 2 is a formal restatement of the patterns using existing definitions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HPC-AI Coupling Methodology for Scientific Applications." pith.science (2026). https://pith.science/paper/LLKEI37M

@misc{pith2026250701025,
  author       = {Pith},
  title        = {Pith review of: HPC-AI Coupling Methodology for Scientific Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LLKEI37M}},
  note         = {Machine review of arXiv:2507.01025}
}
read the original abstract

Artificial intelligence (AI) technologies have fundamentally transformed numerical-based high-performance computing (HPC) applications with data-driven approaches and endeavored to address existing challenges, e.g. high computational intensity, in various scientific domains. In this study, we explore the scenarios of coupling HPC and AI (HPC-AI) in the context of emerging scientific applications, presenting a novel methodology that incorporates three patterns of coupling: surrogate, directive, and coordinate. Each pattern exemplifies a distinct coupling strategy, AI-driven prerequisite, and typical HPC-AI ensembles. Through case studies in materials science, we demonstrate the application and effectiveness of these patterns. The study highlights technical challenges, performance improvements, and implementation details, providing insight into promising perspectives of HPC-AI coupling. The proposed coupling patterns are applicable not only to materials science but also to other scientific domains, offering valuable guidance for future HPC-AI ensembles in scientific discovery.

Figures

Figures reproduced from arXiv: 2507.01025 by the authors.

Figure 1
Figure 1. Illustrating the three patterns of HPC-AI coupling. (1) Surrogate pattern: AI models are used to replace part of entire simulations. (2) Directive pattern: AI provides real-time guidance to HPC simulations, optimizing parameters or providing intermediate feedback. (3) Coordinate pattern: HPC and AI operate interactively, exchanging data and feedback to solve complex problems in tandem. we develop patterns that exemp… view at source ↗
Figure 2
Figure 2. a. Schematic diagram of Crysformer replacing DFT for material property prediction. b. Methodology for 3D graph embedding method. c. Equivariant graph attention layer. principles Geerlings et al. (2003). It provides a balance between accuracy and computational feasibility, making it a cornerstone in the simulation of solid-state systems. How￾ever, DFT remains computationally expensive, especially for large or complex… view at source ↗
Figure 3
Figure 3. a. A projected 3D material structure search space. b. Schematic of the crystal diffusion model. c. Schematic of crystal ab initio generation. d. DFT calculation workflow of the generated material structure. LA,continuous = EϵA∼N(0,I),t∼U(1,T) [∥ϵA − ϵˆA(Mt, t)∥ 2 2 ]. (12) The combined training objective for the joint diffusion model, encompassing L, F, and A, is defined as follows: LM = λLLL + λF LF + λALA. (13) Fo… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Hierarchical ReAct agent planning in coordinate pattern. Deployment via a standardized langchain interface with support for hierarchical tool invocation, including material data repository, DFT calculation workflows and AI pre-trained models. calculation database: Sear…
Figure 5
Figure 5. Figure 5: Impact of dataset size on MAE performance for bandgap property prediction. on highend HPC resources. The AI model training and inference are carried out on NVIDIA A800 GPUs. For DFT calculations, we utilize the VASP version 5.4.4 Wang and Pickett (1983); Chan and Ceder…
Figure 6
Figure 6. Figure 6: Experimental Results. a. Throughput of structure generation using DiffCSP. b. Evolution of the proportion of valid structures over time. calculations with the PBE functional, while being approxi￾mately 886 times more efficient. For the high-accuracy DFT results obtaine…
Figure 7
Figure 7. Figure 7: Statistics of Material Structures Generated by Generative Models. a. Frequency distribution of chemical species. b. Distribution of atom counts in primary unit cells. c. Classification distribution across crystal systems [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Frequency distribution of predicted bandgap a and formation energy values b, respectively. Frequency distribution of DFT calculated bandgap c and formation energy d values, respectively. 4. Structural similarity. A structural similarity algo￾rithm is employed to furthe…
Figure 9
Figure 9. Figure 9: Multi-workflow retrieval-augmented generation for materials informatics [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Structural design for Fe2O3: AI-generated material structures optimized via DFT, resulting in 14 valid structures and their corresponding space groups after multiple iterations. MAEs of 0.081–0.136 eV/atom Xie and Grossman (2018). We conservatively adopt 0.12 eV/atom …
Figure 11
Figure 11. Figure 11: Iterating 50 steps to discover potential novel structures. Failed and duplicate material structures are omitted here. 5.1 Generalization Capacity Surrogate pattern leverages data-driven AI models to approximate HPC computations. Its generalization capacity is limited …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 34 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter doi edition editor eid howpublished institution isbn journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize ":" * " " *...

  3. [3]

    (2024) Accurate structure prediction of biomolecular interactions with alphafold 3

    Abramson J, Adler J, Dunger J, Evans R, Green T, Pritzel A, Ronneberger O, Willmore L, Ballard AJ, Bambrick J et al. (2024) Accurate structure prediction of biomolecular interactions with alphafold 3. Nature : 1--3

    work page 2024

  4. [4]

    (2023) Gpt-4 technical report

    Achiam J, Adler S, Agarwal S, Ahmad L, Akkaya I, Aleman FL, Almeida D, Altenschmidt J, Altman S, Anadkat S et al. (2023) Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  5. [5]

    (2018) Big data and extreme-scale computing: Pathways to convergence-toward a shaping strategy for a future software and data ecosystem for scientific inquiry

    Asch M, Moore T, Badia R, Beck M, Beckman P, Bidot T, Bodin F, Cappello F, Choudhary A, De Supinski B et al. (2018) Big data and extreme-scale computing: Pathways to convergence-toward a shaping strategy for a future software and data ecosystem for scientific inquiry. The International Journal of High Performance Computing Applications 32(4): 435--479

    work page 2018

  6. [6]

    Journal of computational chemistry 28(7): 1177--1186

    Call ST, Zubarev DY and Boldyrev AI (2007) Global minimum structure searches via particle swarm optimization. Journal of computational chemistry 28(7): 1177--1186

    work page 2007

  7. [7]

    Physical review letters 105(19): 196403

    Chan M and Ceder G (2010) Efficient band gap prediction for solids. Physical review letters 105(19): 196403

    work page 2010

  8. [8]

    Chemistry of Materials 31(9): 3564--3572

    Chen C, Ye W, Zuo Y, Zheng C and Ong SP (2019) Graph networks as a universal machine learning framework for molecules and crystals. Chemistry of Materials 31(9): 3564--3572. doi:10.1021/acs.chemmater.9b01294. ://doi.org/10.1021/acs.chemmater.9b01294

    doi:10.1021/acs.chemmater.9b01294 2019

Show all 44 references
  1. [9]

    Nature Computational Science 1(1): 46--53

    Chen C, Zuo Y, Ye W, Li X and Ong SP (2021) Learning properties of ordered and disordered materials from multi-fidelity data. Nature Computational Science 1(1): 46--53

    2021

  2. [10]

    The Journal of Physical Chemistry C 126(38): 16297--16305

    Chen P, Chen J, Yan H, Mo Q, Xu Z, Liu J, Zhang W, Yang Y and Lu Y (2022) Improving material property prediction by leveraging the large-scale computational database and deep learning. The Journal of Physical Chemistry C 126(38): 16297--16305

    2022

  3. [11]

    arXiv preprint arXiv:2401.17244

    Chiang Y, Hsieh E, Chou CH and Riebesell J (2024) Llamp: Large language model made powerful for high-fidelity materials knowledge retrieval and distillation. arXiv preprint arXiv:2401.17244

    2024 arXiv

  4. [12]

    npj Computational Materials 7(1): 1--8

    Choudhary K and DeCost B (2021) Atomistic line graph neural network for improved materials property predictions. npj Computational Materials 7(1): 1--8

    2021

  5. [13]

    Scientific data 5(1): 1--12

    Choudhary K, Zhang Q, Reid AC, Chowdhury S, Van Nguyen N, Trautt Z, Newrock MW, Congo FY and Tavazza F (2018) Computational screening of high-performance optoelectronic materials using optb88vdw and tb-mbj formalisms. Scientific data 5(1): 1--12

    2018

  6. [14]

    In: The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023

    Corso G, St \" a rk H, Jing B, Barzilay R and Jaakkola TS (2023) Diffdock: Diffusion steps, twists, and turns for molecular docking. In: The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net. ://openreview....

    2023

  7. [15]

    Journal of Open Source Software 4(38): 1361

    Davies DW, Butler KT, Jackson AJ, Skelton JM, Morita K and Walsh A (2019) Smact: Semiconducting materials by analogy and chemical theory. Journal of Open Source Software 4(38): 1361

    2019

  8. [16]

    Nature Reviews Physics 6(10): 621--627

    Dongarra J and Keyes D (2024) The co-evolution of computational physics and high-performance computing. Nature Reviews Physics 6(10): 621--627

    2024

  9. [17]

    arXiv preprint arXiv:2003.03123

    Gasteiger J, Gro J and G \"u nnemann S (2020) Directional message passing for molecular graphs. arXiv preprint arXiv:2003.03123

    2020 arXiv

  10. [18]

    Chemical reviews 103(5): 1793--1874

    Geerlings P, De Proft F and Langenaeker W (2003) Conceptual density functional theory. Chemical reviews 103(5): 1793--1874

    2003

  11. [19]

    Computer Physics Communications 175(11): 713--720

    Glass CW, Oganov AR and Hansen N (2006) Uspex—evolutionary crystal structure prediction. Computer Physics Communications 175(11): 713--720. doi:https://doi.org/10.1016/j.cpc.2006.07.020. ://www.sciencedirect.com/science/article/pii/S0010465506002931

    2006 doi

  12. [20]

    In: International Conference on Machine Learning

    Hoogeboom E, Satorras VG, Vignac C and Welling M (2022) Equivariant diffusion for molecule generation in 3d. In: International Conference on Machine Learning. PMLR, pp. 8867--8887

    2022

  13. [21]

    Science 381(6654): 170--175

    Huang B, von Rudorff GF and von Lilienfeld OA (2023) The central role of density functional theory in the ai age. Science 381(6654): 170--175

    2023

  14. [22]

    Nature communications 12(1): 2312

    Jablonka KM, Jothiappan GM, Wang S, Smit B and Yoo B (2021) Bias free multiobjective active learning for materials design and discovery. Nature communications 12(1): 2312

    2021

  15. [23]

    In: SC20: International conference for high performance computing, networking, storage and analysis

    Jia W, Wang H, Chen M, Lu D, Lin L, Car R, Weinan E and Zhang L (2020) Pushing the limit of molecular dynamics with ab initio accuracy to 100 million atoms with machine learning. In: SC20: International conference for high performance computing, networking, storage and analysi...

    2020

  16. [24]

    (2021) Highly accurate protein structure prediction with alphafold

    Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, Tunyasuvunakool K, Bates R, Z \' dek A, Potapenko A et al. (2021) Highly accurate protein structure prediction with alphafold. nature 596(7873): 583--589

    2021

  17. [25]

    Nature Communications 15(1): 4705

    Kang Y and Kim J (2024) Chatmof: an artificial intelligence system for predicting and generating metal-organic frameworks using large language models. Nature Communications 15(1): 4705

    2024

  18. [26]

    npj Computational Materials 10(1): 161

    Koker T, Quigley K, Taw E, Tibbetts K and Li L (2024) Higher-order equivariant neural networks for charge density prediction in materials. npj Computational Materials 10(1): 161

    2024

  19. [27]

    Nature Computational Science 2(6): 367--377

    Li H, Wang Z, Zou N, Ye M, Xu R, Gong X, Duan W and Xu Y (2022) Deep-learning density functional theory hamiltonian for efficient ab initio electronic-structure calculation. Nature Computational Science 2(6): 367--377

    2022

  20. [28]

    In: International Conference on Learning Representations

    Liao YL and Smidt T (2023) Equiformer: Equivariant graph attention transformer for 3d atomistic graphs. In: International Conference on Learning Representations

    2023

  21. [29]

    In: Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024

    Lin P, Chen P, Jiao R, Mo Q, Cen J, Huang W, Liu Y, Huang D and Lu Y (2024) Equivariant diffusion for crystal structure prediction. In: Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 . ://openreview.net/forum?id=VRv8KjJNuj

    2024

  22. [30]

    (2024) Sora: A review on background, technology, limitations, and opportunities of large vision models

    Liu Y, Zhang K, Li Y, Yan Z, Gao C, Chen R, Yuan Z, Huang Y, Sun H, Gao J et al. (2024) Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177

    2024 arXiv

  23. [31]

    Nature Machine Intelligence : 1--11

    M Bran A, Cox S, Schilter O, Baldassari C, White AD and Schwaller P (2024) Augmenting large language models with chemistry tools. Nature Machine Intelligence : 1--11

    2024

  24. [32]

    In: International Conference on Machine Learning

    Nichol AQ and Dhariwal P (2021) Improved denoising diffusion probabilistic models. In: International Conference on Machine Learning. PMLR, pp. 8162--8171

    2021

  25. [33]

    Journal of Chemical Physics 124(24): 201--419

    Oganov AR and Glass CW (2006) Crystal structure prediction using ab initio evolutionary techniques: Principles and applications. Journal of Chemical Physics 124(24): 201--419

    2006

  26. [34]

    Journal of Physics: Condensed Matter 23(5): 053201

    Pickard CJ and Needs R (2011) Ab initio random structure searching. Journal of Physics: Condensed Matter 23(5): 053201

    2011

  27. [35]

    Physical review letters 51(7): 597

    Wang C and Pickett W (1983) Density-functional theory of excitation spectra of semiconductors: application to si. Physical review letters 51(7): 597

    1983

  28. [36]

    ACM Computing Surveys 56(12): 1--38

    Wang H, Yan H, Rong C, Yuan Y, Jiang F, Han Z, Sui H, Jin D and Li Y (2024) Multi-scale simulation of complex systems: a perspective of integrating knowledge and data. ACM Computing Surveys 56(12): 1--38

    2024

  29. [37]

    Advances in Neural Information Processing Systems 31

    Weiler M, Geiger M, Welling M, Boomsma W and Cohen TS (2018) 3d steerable cnns: Learning rotationally equivariant features in volumetric data. Advances in Neural Information Processing Systems 31

    2018

  30. [38]

    Nature materials 7(12): 937--946

    Woodley SM and Catlow R (2008) Crystal structure prediction from first principles. Nature materials 7(12): 937--946

    2008

  31. [39]

    Physical review letters 120(14): 145301

    Xie T and Grossman JC (2018) Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties. Physical review letters 120(14): 145301

    2018

  32. [40]

    ArXiv abs/2209.11807

    Yan K, Liu Y, Lin YC and Ji S (2022) Periodic graph transformers for crystal material property prediction. ArXiv abs/2209.11807

    2022 arXiv

  33. [41]

    Nature Medicine : 1--11

    Yang Y, Zhang H, Gichoya JW, Katabi D and Ghassemi M (2024) The limits of fair medical imaging ai in real-world generalization. Nature Medicine : 1--11

    2024

  34. [42]

    (2023) Dpa-2: Towards a universal large atomic model for molecular and material simulation

    Zhang D, Liu X, Zhang X, Zhang C, Cai C, Bi H, Du Y, Qin X, Huang J, Li B et al. (2023) Dpa-2: Towards a universal large atomic model for molecular and material simulation. arXiv preprint arXiv:2312.15492

    2023 arXiv

  35. [43]

    , " * write output.state after.block = add.period write newline

    ENTRY address archive author booktitle chapter doi edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type url volume year label INTEGERS output.state before.all mid.sentence after.sentence aft...

  36. [44]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.