Pith. sign in

REVIEW 4 major objections 4 minor 23 references

Temperature-Driven Sequential Modeling for the Prediction of Annual Power Conversion Efficiency Profiles of Organic Photovoltaic Materials: Douala Case Study

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read Molecular motion reshapes organic solar efficiency forecasts

desk verdict A promising climate-aware screening pipeline undermined by target leakage in the temporal test; the GNN surrogate is solid but the load-bearing claim about thermal dynamics does not hold as written. read the letter →

arxiv 2608.11261 v1 pith:M5QMNMXN submitted 2026-08-09 cond-mat.mtrl-sci cs.LGphysics.chem-ph

classification cond-mat.mtrl-scics.LGphysics.chem-ph
keywords organicphotovoltaicspowerconversionefficiencymoleculardynamicsGFN2-xTBequivariantgraphneuralnetworksequentialdeeplearningseasonalstabilityscoretropicalclimatedeployment
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

This paper argues that room-temperature static screening of organic photovoltaic (OPV) materials mispredicts real-world performance in tropical climates, and that modeling how molecules vibrate and twist under heat yields materially better annual efficiency forecasts. It builds a two-stage pipeline: an equivariant graph neural network trained on thousands of thermally distorted molecular dynamics snapshots replaces expensive quantum chemistry, and sequential deep learning models then forecast a 52-week power conversion efficiency (PCE) profile from climate data for Douala, Cameroon. The central positive claim is that sequential models trained on full molecular dynamics trajectories beat time-averaged static baselines by 35% to 48% relative MAE, which the authors read as evidence that thermal conformational dynamics carry information beyond mean geometry. If correct, deployment decisions—which molecules to actually put on tropical rooftops—should be made from climate-conditioned annual profiles and a seasonal stability score, not from standard testing condition (STC) PCE alone.

What carries the argument

The load-bearing mechanism is a two-level surrogate stack. Level one is a PaiNN equivariant graph neural network trained on ~120,600 GFN2-xTB molecular dynamics snapshots of 268 Neyman-stratified Clean Energy Project molecules; it predicts HOMO/LUMO energies, reorganization energies, and Scharber PCE at roughly a 1050x speedup over explicit GFN2-xTB. Level two is a sequential deep learning head (Transformer, LSTM, GRU) fed a 52-week multivariate time series of climate variables and predicted electronic properties, producing the annual PCE profile. The NOCT thermal balance model converts ambient temperature and irradiance into panel cell temperature, and a linear calibration maps tight-binding orbital energies onto the CEP DFT reference scale.

What would settle it

Retrain the sequential model without the instantaneous PCE(t) feature while keeping all other trajectory features; if the relative MAE improvement over the static baseline drops to near zero or becomes insignificant, the paper's evidence that thermal conformational dynamics carry information beyond mean geometry is not supported. A secondary check would shuffle the temporal order of MD snapshots within each molecule and confirm that the sequential model's advantage disappears when the actual dynamics are destroyed.

Watch

Extended reading notes

Core claim

The paper claims that finite-temperature conformational dynamics encode physically relevant signal for OPV efficiency forecasting that static single-point calculations miss. Concretely, it reports that sequential forecasting models (Transformer, LSTM, GRU) trained on full multimodality time series, including cell temperature, irradiance, humidity, wind speed, orbital energies, and instantaneous Scharber PCE, achieve 35% to 48% relative MAE improvement over a collapsed static mean baseline on annual PCE prediction. The authors further claim that this climate-native forecast outperforms the traditional static Scharber model when validated against 350 experimentally measured HOPV15 devices, with higher coefficient of determination and lower error, and they introduce a clamped seasonal stability score that reranks donor molecules by performance consistency under tropical microclimates. The paper's stated conclusion is that static STC screening systematically overestimates real-world energy yield for thermally sensitive donor structures, and that climate-conditioned trajectory-based screening is a scalable alternative.

Load-bearing premise

The claim that thermal dynamics carry information beyond mean geometry depends on the sequential model actually learning from trajectory content, but because each input time step already contains an instantaneous Scharber PCE derived from the same electronic properties as the target PCE series, the model could in principle copy the target from its inputs rather than learn from molecular motion.

Editorial extensions

If this is right

  • If the central claim holds, static STC-based virtual screening should be supplemented by climate-conditioned annual PCE profiles for any OPV deployment in hot, seasonally varying climates.
  • The seasonal stability score provides an operational ranking criterion distinct from peak STC PCE; molecules with high STC PCE but large seasonal PCE swings would be deprioritized for tropical deployment.
  • The reported 35–48% MAE improvement implies that including trajectory-level conformational dynamics could materially reduce the error of high-throughput OPV screening pipelines without requiring explicit quantum chemistry at every snapshot.
  • Zero-shot transfer of the framework to other Cameroonian cities suggests that, within the trained temperature envelope, the same models may forecast PCE profiles for other tropical locations using only local climate data.

Reading between the lines

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

  • The paper's temporal-information claim is not yet sealed: because the input vector at each time step includes an instantaneous Scharber PCE computed from the very electronic properties that define the target annual PCE series, part or all of the sequential model's edge over static baselines could come from copying the target from its own inputs. A cleaner test would remove PCE(t) from the feature
  • A natural extension the authors do not pursue is using the same pipeline to predict not only mean annual PCE but also degradation-rate distributions, since their own Marcus–Scharber correlation links reorganization energy to seasonal PCE drop—this could feed into levelized cost of energy models for tropical solar farms.
  • The framework's physics is largely in vacuum or implicit-solvent MD; bridging to solid-state morphology, which the paper itself flags as a failure regime for highly pi-stacking donors, would likely improve the 4.2% of large-outlier cases and is a testable next step.
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. The paper proposes a 'Climate-Native' computational framework for predicting annual power conversion efficiency (PCE) profiles of organic photovoltaic (OPV) donor molecules under tropical operating conditions (Douala, Cameroon). The pipeline combines GFN2-xTB molecular dynamics, an equivariant graph neural network (PaiNN) surrogate for electronic properties, and sequential deep learning models (LSTM, GRU, Transformer) trained on 52-week time series built from NASA POWER climate data. The authors report that sequential models outperform time-averaged static baselines by 35%–48% relative MAE, claim experimental validation on 350 HOPV15 devices, introduce a seasonal stability score for reranking candidates, and demonstrate zero-shot transfer to Yaoundé and Maroua. The central scientific claim is that thermal conformational dynamics carry predictive information beyond mean molecular geometry.

Significance. If established, the central claim would be a meaningful step toward climate-aware virtual screening of OPV materials, with practical relevance for deployment in tropical regions. The GNN surrogate benchmark is a plausible and potentially useful contribution: it uses a group-level molecular split to avoid leakage, compares several architectures, and reports a large inference speedup. The seasonal stability score is a reasonable screening metric. However, the headline temporal-information claim is currently compromised by a feature-asymmetry problem: the sequential models receive per-timestep electronic properties (and possibly instantaneous Scharber PCE) from which the target PCE can be reconstructed by the known Scharber equations, while the collapsed baseline does not receive this information. Therefore the reported improvements do not, as presented, demonstrate that thermal dynamics encode signal beyond mean geometry. The HOPV15 validation also compares annual operational PCE with experimental STC efficiencies, an apples-to-oranges comparison. These issues affect the paper's main conclusions.

major comments (4)
  1. [Section 2.2, Check 3 and Table 2] The temporal information test does not control for feature asymmetry. The sequential model receives per-timestep electronic features (EHOMO(t), Egap(t), and per Section 3.5 instantaneous Scharber PCE), from which the weekly PCE target is directly computed via the Scharber equations (6)–(8). The 'Collapsed Static Mean' baseline receives only time-averaged vectors and thus cannot exploit the per-timestep values. The reported 38.4% MAE improvement in Check 3 and the 35%–48% improvements in Table 2 could therefore reflect the model's ability to reproduce a deterministic mapping from inputs to target rather than any physical information carried by conformational dynamics. Please provide a leakage-free ablation—for example, giving the same per-timestep features to the baseline (e.g., by averaging the model outputs over the week) or forecasting a genuinely future value—before the temporal-information claim can be accepted.
  2. [Section 2.4] The HOPV15 validation compares the Climate-Native annual forecast (μ_PCE under Douala operating conditions) with experimental PCEs that are measured under standard illumination conditions. This is not a like-for-like comparison: the static STC Scharber model is the appropriate baseline for STC-measured efficiencies, and the reported improvement in R² (0.78 vs 0.54) could arise from differences other than climate-native modeling, such as the linear calibration to the CEP DFT scale or the learned sequential mapping. Please either compare STC predictions from the same pipeline or justify why μ_PCE is the correct predictor of STC device measurements.
  3. [Section 3.5 vs Section 2.2] The composition of the input vector is described inconsistently. Section 2.2 lists x(t) = [T_cell(t), G(t), RH2M(t), WS2M(t), EHOMO(t), Egap(t), PCE_STC]^T, while Section 3.5 states that input features include 'instantaneous Scharber PCE.' If the latter is used, the model has direct access to the target variable (or a near copy of it) at each timestep, which would be direct target leakage. Please clarify which input set was actually used and, if instantaneous PCE is an input, explain how the temporal-information claim survives its removal.
  4. [Abstract and Conclusions] The statement that sequential models 'confirm that thermal conformational dynamics carry information beyond mean geometry' is the paper's central claim, but it is not supported by the current experimental design, which contrasts models with and without time-resolved inputs rather than models that do and do not see conformational dynamics. At minimum, the claim should be softened to 'sequential processing of time-resolved electronic properties improves forecasting' unless the proposed leakage-free ablation demonstrates otherwise.
minor comments (4)
  1. [Figure 5b] The relative MAE improvements reported in Figure 5b (+25.0%, +37.5%, +43.8%, +46.9%) do not exactly match the values in Table 2 (e.g., LSTM 43.3%, Transformer 45.5%). Please reconcile the two presentations.
  2. [Equation (1)] The linear calibration E_DFT = 1.12 × E_xTB − 0.45 eV is presented with R² = 0.92 and MAE = 0.048 eV; it would be helpful to state explicitly how the parameters were fitted and whether the calibration was applied to the GNN outputs before computing the surrogate error metrics in Table 1.
  3. [Section 2.5] The Sobol sensitivity analysis mentioned in the text is not described in Methods or in the Supplementary Information; please provide the procedure, the input distributions, and the resulting indices, or remove the claim.
  4. [Section 3.4, Eq. (8)] The instantaneous PCE is computed from GNN-predicted orbital energies, which have MAEs of ~0.03 eV. Propagating these uncertainties into the final PCE forecasts would strengthen the quantitative claims and is currently omitted.

Circularity Check

1 steps flagged · score 7.0 of 10

Temporal-information claim is compromised by target leakage: sequential inputs include instantaneous Scharber PCE, which is the same quantity used to define the weekly PCE forecast target.

  1. self definitional [Section 2.2 Check 3 and Section 3.5 (GNN Surrogate and Sequential Forecasting Models)]
    "Input feature vectors at timestep t comprised: internal coordinate BAT dense embeddings, GNN-predicted electronic properties (E_HOMO, E_LUMO, E_gap, λ±_vert), instantaneous Scharber PCE, and continuous panel temperature T_cell(t). Models were trained using AdamW (10−4 learning rate, early stopping on validation loss) to predict weekly PCE forecasts over the 52-week annual sequence."

    The weekly PCE forecast target is generated by the corrected Scharber engine (Eqs. 6-8) from the same instantaneous electronic properties (E_HOMO, E_gap) that also define the instantaneous Scharber PCE input channel. A sequential model can approximate the weekly target by temporally aggregating the instantaneous PCE channel (e.g., by averaging over the week), while the 'Collapsed Static Mean' baseline in Table 2 is not given this temporal PCE channel. The reported 38.4% improvement in Check 3 and the 35-48% improvements in Table 2 therefore reduce to feature asymmetry rather than evidence that thermal conformational dynamics carry information beyond mean geometry.

full rationale

The paper's headline claim that sequential models on full MD trajectories outperform static baselines because thermal dynamics encode information beyond mean geometry rests on Check 3 and Table 2. That test is circular: the sequential input vector includes instantaneous Scharber PCE, a quantity computed by the same Scharber equations and from the same orbital properties that define the weekly PCE forecast target. The collapsed static baseline lacks this channel, so the improvement is largely explained by the model copying or averaging the target-derived input. The GNN surrogate validation, HOPV15 external comparison, and zero-shot transfer are not circular in themselves, and the linear calibration in Eq. 1 is an explicit fit rather than a hidden circular step. However, the central temporal-information result is load-bearing for the paper's conclusion that static STC screening systematically overestimates real-world yield, and that result is compromised by the input/target overlap. Hence a score of 7 is appropriate: the central claim partially reduces to a self-definitional input feature.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claims rest on a chain of modeling choices: the linear calibration of GFN2-xTB to CEP DFT, the Scharber model's fixed device constants, the NOCT thermal model, gas-phase MD as a proxy for film behavior, and the comparability of Douala annual mean PCE to STC experimental PCE. None of these are independently validated in this paper, and the most fragile is the condition matching for HOPV15.

free parameters (4)
  • Linear calibration slope for orbital energies = 1.12
    Eq. 1 maps GFN2-xTB orbital energies to CEP DFT scale; fitted to CEP BP86 reference data with R2=0.92 and MAE=0.048 eV. Every downstream PCE value depends on this fit.
  • Linear calibration intercept for orbital energies = -0.45 eV
    Same Eq. 1 fit; absorbs uniform solvation and method shifts. Not independently validated.
  • NOCT panel temperature parameter = 45.0 degrees Celsius
    Used in Eq. 4 to convert ambient temperature and irradiance to cell temperature. This is a chosen device parameter, not fitted here, and affects the entire thermal profile.
  • OOD uncertainty inflation factor kappa = 1.85
    Hand-set in the Maroua out-of-distribution fallback handler (Section 2.6); no justification or calibration is provided.
assumptions (5)
  • domain assumption Scharber device model with fixed FF=0.65, EQE=0.65, Voc loss 0.3 V, and acceptor LUMO levels of -4.3 eV or -3.9 eV
    Equations 6-8 in Section 3.4 convert orbital energies to PCE using fixed device parameters from ref. [3]. No uncertainty is propagated and no experimental calibration is applied to these constants.
  • domain assumption Gas-phase GFN2-xTB MD with ALPB toluene solvation captures the conformational ensemble relevant to bulk-heterojunction devices
    Section 3.3-3.4 uses single-molecule vacuum trajectories and a low-dielectric solvation proxy. The paper itself lists morphology, steric barriers, and non-radiative recombination as outlier failure regimes in Section 2.4.
  • domain assumption CEP BP86/def2-SVP DFT reference values are valid ground truth for orbital energy calibration
    Eq. 1 is fit to the CEP DFT reference. The paper acknowledges BP86 self-interaction error but still uses this reference as the acceptance baseline for the GNN surrogate.
  • domain assumption Annual mean PCE under Douala operating conditions is comparable to HOPV15 experimental PCEs measured at standard testing conditions
    Section 2.4 compares climatological annual mean PCE to experimental STC device PCEs. If the conditions differ, the reported R2 improvement is not a clean validation.
  • domain assumption NASA POWER weather data and the linear NOCT heat balance model represent actual panel operating temperatures
    Section 3.2 uses Eq. 4 with a fixed NOCT value. This is a standard engineering approximation but is not validated against field panel temperature measurements.
invented entities (1)
  • Clamped seasonal stability score S_stability
    purpose: Rerank OPV donor molecules by performance consistency under tropical seasonal temperature variation
    Defined in Eq. 2 as max(0, 1 - sigma_PCE/(mu_PCE + 1e-6)). It is a normalized coefficient of variation derived entirely from the model's own PCE time series, with no external benchmark or experimental validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Temperature-Driven Sequential Modeling for the Prediction of Annual Power Conversion Efficiency Profiles of Organic Photovoltaic Materials: Douala Case Study." pith.science (2026). https://pith.science/paper/M5QMNMXN

@misc{pith2026260811261,
  author       = {Pith},
  title        = {Pith review of: Temperature-Driven Sequential Modeling for the Prediction of Annual Power Conversion Efficiency Profiles of Organic Photovoltaic Materials: Douala Case Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M5QMNMXN}},
  note         = {Machine review of arXiv:2608.11261}
}
abstract

Organic photovoltaic (OPV) materials are promising candidates for distributed solar energy in tropical regions, yet existing virtual screening tools report static power conversion efficiency (PCE) values at standard testing conditions (STC) that fail to capture the temperature-driven performance degradation experienced under real deployment conditions. Here we introduce a Climate-Native computational framework that forecasts the annual PCE profile of OPV donor molecules under geographically realistic operating conditions. The framework combines GFN2-xTB molecular dynamics with an equivariant graph neural network surrogate ($268$ Neyman-stratified CEP molecules; $120,600$ training geometries; $\sim 1050\times$ speedup over explicit quantum chemistry) and sequential deep learning models trained on annual time series anchored in NASA POWER climate data for Douala, Cameroon, and validated by zero-shot transfer to Yaound\'e and Maroua. Applied to $\sim 30,000$ molecules from the Harvard Clean Energy Project (CEP) and validated against $350$ HOPV15 experimental device measurements, the framework demonstrates that sequential models trained on full molecular dynamics trajectories outperform time-averaged baselines ($35\%$-$48\%$ relative MAE improvement over static baselines), confirming that thermal conformational dynamics carry information beyond mean geometry. We further introduce a seasonal stability score that reranks OPV candidates by performance consistency under tropical conditions, identifying molecules whose deployment suitability differs substantially from their static PCE ranking.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 11 canonical work pages

  1. [1]

    Organics5(4), 520–560 (2024) https://doi

    Nowsherwan, G.A., Ali, Q., Ali, U.F., Ahmad, M., Khan, M., Hussain, S.S.: Advances in Organic Materials for Next-Generation Optoelectronics: Potential and Challenges. Organics5(4), 520–560 (2024) https://doi. org/10.3390/org5040028 . Accessed 2025-11-20

  2. [2]

    Hachmann, J., Olivares-Amaya, R., Atahan-Evrenk, S., Amador-Bedolla, C., S´ anchez-Carrera, R.S., Gold- Parker, A., Vogt, L., Brockway, A.M., Aspuru-Guzik, A.: The Harvard Clean Energy Project: Large-Scale Computational Screening and Design of Organic Photovoltaics on the World Community Grid. J. Phys. Chem. Lett.2(17), 2241–2251 (2011) https://doi.org/10...

  3. [3]

    Scharber, M.C., M¨ uhlbacher, D., Koppe, M., Denk, P., Waldauf, C., Heeger, A.J., Brabec, C.J.: Design Rules for Donors in Bulk-Heterojunction Solar Cells—Towards 10% Energy-Conversion Efficiency. Adv. Mater. 18(6), 789–794 (2006) https://doi.org/10.1002/adma.200501717

  4. [4]

    Zhugayevych, A., Tretiak, S.: Theoretical Description of Structural and Electronic Properties of Organic Photovoltaic Materials. Annu. Rev. Phys. Chem.66(1), 305–330 (2015) https://doi.org/10.1146/ annurev-physchem-040214-121440 . Accessed 2025-11-20

  5. [5]

    J Cheminform15(1), 69 (2023) https://doi.org/10.1186/s13321-023-00737-5

    Panwar, P., Yang, Q., Martini, A.: PyL3dMD: Python LAMMPS 3D molecular descriptors package. J Cheminform15(1), 69 (2023) https://doi.org/10.1186/s13321-023-00737-5 . Accessed 2025-11-25

  6. [6]

    Schleder, G.R., Padilha, A.C.M., Acosta, C.M., Costa, M., Fazzio, A.: From DFT to machine learning: recent approaches to materials science–a review. J. Phys. Mater.2(3), 032001 (2019) https://doi.org/10. 1088/2515-7639/ab084b . Accessed 2025-11-25

  7. [7]

    npj Comput Mater3(1), 54 (2017) https://doi.org/10.1038/ s41524-017-0056-5

    Ramprasad, R., Batra, R., Pilania, G., Mannodi-Kanakkithodi, A., Kim, C.: Machine learning in materials informatics: recent applications and prospects. npj Comput Mater3(1), 54 (2017) https://doi.org/10.1038/ s41524-017-0056-5 . Accessed 2025-11-26

  8. [8]

    npj Comput Mater7(1), 185 (2021) https://doi.org/10.1038/s41524-021-00650-1

    Choudhary, K., DeCost, B.: Atomistic Line Graph Neural Network for improved materials property predictions. npj Comput Mater7(1), 185 (2021) https://doi.org/10.1038/s41524-021-00650-1 . Accessed 2025-11-25

Show all 23 references
  1. [9]

    Chen, C., Ye, W., Zuo, Y., Zheng, C., Ong, S.P.: Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals. Chem. Mater.31(9), 3564–3572 (2019) https://doi.org/10.1021/acs.chemmater. 9b01294 . Accessed 2025-11-25

  2. [10]

    Sch¨ utt, K.T., Sauceda, H.E., Kindermans, P.-J., Tkatchenko, A., M¨ uller, K.-R.: SchNet – A deep learning architecture for molecules and materials. J. Chem. Phys.148(24), 241722 (2018) https://doi.org/10.1063/1. 5019779 11

  3. [11]

    In: Proceedings of the 38th International Conference on Machine Learning (ICML 2021), pp

    Sch¨ utt, K., Unke, O., Gastegger, M.: Equivariant message passing for the prediction of tensorial properties and molecular spectra. In: Proceedings of the 38th International Conference on Machine Learning (ICML 2021), pp. 9377–9388. PMLR, ??? (2021).https://proceedings.mlr.pr...

  4. [12]

    In: Advances in Neural Information Processing Systems 35 (NeurIPS 2022) (2022)

    Batatia, I., Kovacs, D.P., Simm, G.N.C., Ortner, C., Cs´ anyi, G.: MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields. In: Advances in Neural Information Processing Systems 35 (NeurIPS 2022) (2022). arXiv:2206.07697.https://arxiv....

  5. [13]

    Neural Comput.9(8), 1735–1780 (1997) https: //doi.org/10.1162/neco.1997.9.8.1735

    Hochreiter, S., Schmidhuber, J.: Long Short-Term Memory. Neural Comput.9(8), 1735–1780 (1997) https: //doi.org/10.1162/neco.1997.9.8.1735

  6. [14]

    In: Advances in Neural Information Processing Systems 30 (NeurIPS 2017) (2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention Is All You Need. In: Advances in Neural Information Processing Systems 30 (NeurIPS 2017) (2017). arXiv:1706.03762.https://arxiv.org/abs/1706.03762

  7. [15]

    Bannwarth, C., Ehlert, S., Grimme, S.: GFN2-xTB—An Accurate and Broadly Parametrized Self-Consistent Tight-Binding Quantum Chemical Method with Multipole Electrostatics and Density-Dependent Dispersion Contributions. J. Chem. Theory Comput.15(3), 1652–1671 (2019) https://doi.o...

  8. [16]

    In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp

    Cho, K., Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y.: Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation. In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processin...

  9. [17]

    Stackhouse, P.W., Zhang, T., Westberg, D., Barnett, A.J., Bristow, T., Macpherson, B., Hoell, J.M.: The NASA POWER Project: Transitioning Research to Operations. Bull. Am. Meteorol. Soc.100(12), 2379–2396 (2019) https://doi.org/10.1175/BAMS-D-18-0148.1

  10. [18]

    Sci Data3(1), 160086 (2016) https://doi.org/10.1038/sdata

    Lopez, S.A., Pyzer-Knapp, E.O., Simm, G.N., Lutzow, T., Li, K., Seress, L.R., Hachmann, J., Aspuru-Guzik, A.: The Harvard organic photovoltaic dataset. Sci Data3(1), 160086 (2016) https://doi.org/10.1038/sdata. 2016.86 . Accessed 2025-05-18

  11. [19]

    In: Machine Learning for Molecules Workshop, NeurIPS 2020 (2020)

    Gasteiger, J., Giri, S., Margraf, J.T., G¨ unnemann, S.: Fast and Uncertainty-Aware Directional Message Passing for Non-Equilibrium Molecules. In: Machine Learning for Molecules Workshop, NeurIPS 2020 (2020). arXiv:2011.14115.https://arxiv.org/abs/2011.14115

  12. [20]

    Sch¨ utt, K.T., Hessmann, S.S.P., Gebauer, N.W.A., Lederer, J., Gastegger, M.: SchNetPack 2.0: A neural network toolbox for atomistic machine learning. J. Chem. Phys.158(14), 144801 (2023) https://doi.org/10. 1063/5.0138367

  13. [21]

    Kluwer Academic Publishers, Dordrecht (1996)

    Singh, S.: Advanced Sampling Theory with Applications: How Sampling Works. Kluwer Academic Publishers, Dordrecht (1996). https://doi.org/10.1007/978-94-007-0789-4

  14. [22]

    Ehlert, S., Stahn, M., Spicher, S., Grimme, S.: Robust and Efficient Implicit Solvation Model for Fast Semiempirical Methods. J. Chem. Theory Comput.17(7), 4250–4261 (2021) https://doi.org/10.1021/acs. jctc.1c00471

  15. [23]

    Hourahine, B., Aradi, B., Blum, V., Bonaf´ e, F., Buccheri, A., Camacho, C., Cevallos, C., Deshaye, M.Y., Dumitric˘ a, T., Dominguez, A., Ehlert, S., Elstner, M., Heide, T., Hermann, J., Irle, S., Kranz, J.J., K¨ ohler, C., Kowalczyk, T., Kubaˇ r, T., Lee, I.S., Lutsker, V., M...

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.