pith. sign in

arxiv: 2603.15411 · v2 · pith:TTT7SLBDnew · submitted 2026-03-16 · 💻 cs.AI · cs.LG

A Hybrid Modeling Framework for Crop Prediction Tasks via Dynamic Parameter Calibration and Multi-Task Learning

William Solow , Paola Pesantez-Cabrera , Markus Keller , Lav Khot , Sandhya Saisubramanian , Alan Fern This is my paper

Pith reviewed 2026-05-21 10:33 UTC · model grok-4.3

classification 💻 cs.AI cs.LG
keywords hybrid modelingcrop predictionbiophysical modelsmulti-task learningneural networksphenologycold hardinessdifferentiable models
0
0 comments X

The pith

A neural network predicts the parameters of a biophysical crop model to raise accuracy by 60 percent for phenology while preserving biological realism.

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

The paper introduces a hybrid method for forecasting crop states such as phenology stages and cold hardiness. A neural network learns to output the internal parameters of an existing biophysical model rather than directly regressing the target states. Multi-task learning lets the network share information across different crop cultivars even when data for each one is scarce. Because the biophysical model remains in the loop and is kept differentiable, the predictions stay inside biologically valid bounds that pure deep-learning models often break. The authors report that this yields substantially higher accuracy than the biophysical models already used in practice.

Core claim

By training a neural network to predict the parameters of a differentiable biophysical model and training it with multi-task learning across cultivars, the hybrid framework produces more accurate forecasts of phenology and cold hardiness than the deployed biophysical models while automatically satisfying the model’s biological constraints.

What carries the argument

A neural network that outputs the time-varying parameters of a differentiable biophysical model, trained jointly across cultivars via multi-task learning.

If this is right

  • Phenology predictions improve by roughly 60 percent and cold-hardiness predictions by 40 percent relative to current deployed models.
  • Biological constraints remain satisfied without extra penalty terms because the biophysical model itself is never bypassed.
  • Multi-task learning reduces data requirements by letting information flow between cultivars that share similar parameter structures.
  • The same trained network can be used for season-long forecasts that support irrigation, fertilization, and canopy-management decisions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The differentiability requirement opens the door to gradient-based calibration of any biophysical model that can be rewritten in that form.
  • The approach could be tested on other crops or environmental variables where a mechanistic model already exists but needs site-specific tuning.
  • Because parameters are predicted rather than learned as fixed values, the model may adapt more readily when weather or management practices change within a season.

Load-bearing premise

The biophysical model can be written in differentiable form so that a neural network can predict its parameters and the resulting predictions automatically obey biological constraints.

What would settle it

On a held-out real-world dataset for a new cultivar, the hybrid model shows no accuracy gain over the original biophysical model or produces predictions that violate known biological limits such as negative growth rates.

Figures

Figures reproduced from arXiv: 2603.15411 by Alan Fern, Lav Khot, Markus Keller, Paola Pesantez-Cabrera, Sandhya Saisubramanian, William Solow.

Figure 1
Figure 1. Figure 1: Overview of our proposed method using phenology prediction as an example. In this case, seasonal phenological stages guide [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Network architecture of our approach DMC-MTL. The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: In-Season Adaptation with DMC-MTL. Cultivar id [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: The performance of DMC-MTL models compared to deep [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance of the base DMC-MTL with increasing per [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Percentage of all cultivars with cumulative error below [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Daily crop state observations for five cultivars of (a) grape phenology and (b) grape cold-hardiness during a single growing season. [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The distribution of per-stage prediction error (RMSE [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: The performance of DMC-MTL with a sliding weather [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The integrated gradients values of the Base Tempera [PITH_FULL_IMAGE:figures/full_fig_p017_13.png] view at source ↗
read the original abstract

Accurate prediction of crop states (e.g., phenology stages and cold hardiness) is essential for timely farm management decisions such as irrigation, fertilization, and canopy management to optimize crop yield and quality. While traditional biophysical models can be used for season-long predictions, they lack the precision required for site-specific management. Deep learning methods are a compelling alternative, but can produce biologically unrealistic predictions and require large-scale data. We propose a \emph{hybrid modeling} approach that uses a neural network to parameterize a differentiable biophysical model and leverages multi-task learning for efficient data sharing across crop cultivars in data limited settings. By predicting the \emph{parameters} of the biophysical model, our approach improves the prediction accuracy while preserving biological realism. Empirical evaluation using real-world and synthetic datasets demonstrates that our method improves prediction accuracy by 60\% for phenology and 40\% for cold hardiness compared to deployed biophysical models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a hybrid modeling framework for crop state prediction (phenology stages and cold hardiness) that uses a neural network to predict the parameters of a differentiable biophysical model, combined with multi-task learning to share information across cultivars in data-limited regimes. The central claim is that this parameterization improves prediction accuracy (reported as 60% for phenology and 40% for cold hardiness) over deployed biophysical models while preserving biological realism, with supporting experiments on real-world and synthetic datasets.

Significance. If the reported gains can be shown to arise from the hybrid structure rather than calibration alone, the work would offer a practical route to site-specific, biologically constrained forecasts that improve upon both pure mechanistic models and unconstrained deep learning. The emphasis on differentiability and cross-cultivar sharing is a timely contribution to precision agriculture modeling.

major comments (3)
  1. [Abstract and §4.2] Abstract and §4.2 (Experimental Setup): The headline accuracy improvements are stated without specifying the precise error metrics, the data splits used for training versus testing, or whether the baseline biophysical models received equivalent per-site or per-cultivar calibration on the same data. This omission makes it impossible to determine whether the 60% and 40% gains reflect the hybrid architecture, differentiability, or multi-task learning rather than data-driven parameter fitting.
  2. [§3.1] §3.1 (Model Formulation): The framework assumes the biophysical model can be expressed in fully differentiable form so that gradients flow through the neural parameter predictor. The manuscript does not detail how discrete or non-smooth operations (e.g., stage thresholds in phenology or temperature-based hardiness transitions) are handled or approximated, which is load-bearing for the claim of automatic biological constraint enforcement.
  3. [§5] §5 (Results): No ablation experiments isolate the contribution of multi-task learning across cultivars from single-task dynamic calibration, nor do they compare against a version that simply fits the biophysical parameters directly from data without the neural network. Such controls are required to substantiate that the hybrid design itself drives the reported improvements.
minor comments (2)
  1. [Table 1] Table 1: The list of biophysical parameters would benefit from an additional column indicating which parameters are cultivar-specific versus shared.
  2. [Figure 2] Figure 2 caption: Clarify whether the synthetic data generation process re-uses the same biophysical equations as the model being calibrated.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thoughtful and constructive comments, which help clarify the presentation of our hybrid modeling framework. We address each major point below and commit to revisions that strengthen the manuscript's clarity and rigor without altering the core contributions.

read point-by-point responses

  1. Referee: [Abstract and §4.2] The headline accuracy improvements are stated without specifying the precise error metrics, the data splits used for training versus testing, or whether the baseline biophysical models received equivalent per-site or per-cultivar calibration on the same data. This omission makes it impossible to determine whether the 60% and 40% gains reflect the hybrid architecture, differentiability, or multi-task learning rather than data-driven parameter fitting.

    Authors: We agree that these experimental details are essential for interpreting the reported gains. In the revised manuscript, we will explicitly define the error metrics (root mean squared error for phenology stage predictions and mean absolute error for cold hardiness), describe the data partitioning (temporal hold-out by growing season with cross-validation across sites and cultivars), and confirm that baseline biophysical models were calibrated on identical training data using the same per-cultivar and per-site optimization procedure. These additions will appear in the abstract, §4.2, and a new subsection of §5 to better isolate the hybrid and multi-task contributions. revision: yes

  2. Referee: [§3.1] The framework assumes the biophysical model can be expressed in fully differentiable form so that gradients flow through the neural parameter predictor. The manuscript does not detail how discrete or non-smooth operations (e.g., stage thresholds in phenology or temperature-based hardiness transitions) are handled or approximated, which is load-bearing for the claim of automatic biological constraint enforcement.

    Authors: We acknowledge the need for explicit description of differentiability. The biophysical model employs smooth sigmoid-based approximations for discrete stage thresholds and temperature transitions to enable end-to-end gradient flow while preserving the original biological logic. We will expand §3.1 with the precise functional forms of these approximations, their hyperparameters, and a brief analysis of approximation error relative to the non-smooth original, thereby supporting the claim of enforced biological constraints. revision: yes

  3. Referee: [§5] No ablation experiments isolate the contribution of multi-task learning across cultivars from single-task dynamic calibration, nor do they compare against a version that simply fits the biophysical parameters directly from data without the neural network. Such controls are required to substantiate that the hybrid design itself drives the reported improvements.

    Authors: We agree that targeted ablations are required to attribute improvements correctly. The revised §5 will include new experiments comparing (i) the full hybrid multi-task model, (ii) single-task dynamic calibration (no cross-cultivar sharing), and (iii) direct least-squares fitting of biophysical parameters without the neural predictor. Results will be reported on both real-world and synthetic datasets to demonstrate the incremental value of the neural parameterization and multi-task learning. revision: yes

Circularity Check

1 steps flagged

Accuracy gains reduce to data-driven parameter calibration rather than hybrid structure

specific steps
  1. fitted input called prediction [Abstract]
    "By predicting the parameters of the biophysical model, our approach improves the prediction accuracy while preserving biological realism. Empirical evaluation using real-world and synthetic datasets demonstrates that our method improves prediction accuracy by 60% for phenology and 40% for cold hardiness compared to deployed biophysical models."

    The neural network directly outputs the parameters that are then inserted into the biophysical model to generate the phenology and cold-hardiness predictions. The accuracy metric is computed on those outputs. Because parameter prediction is the fitting step itself, any measured improvement over uncalibrated deployed models is equivalent to the calibration effect by construction; the hybrid claim adds no additional independent content to the reported percentages.

full rationale

The paper's central claim is that predicting biophysical model parameters via neural network yields 60% and 40% accuracy improvements over deployed models while preserving realism. However, this improvement is measured on the same crop-state outputs that the fitted parameters directly control. No evidence is provided that baselines received equivalent per-site or per-cultivar calibration, so the reported lift is statistically forced by the act of fitting parameters to data rather than by differentiability, multi-task sharing, or hybrid architecture. This matches the fitted-input-called-prediction pattern with no independent derivation chain shown.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on the assumption that biophysical models admit a differentiable formulation and that multi-task learning across cultivars transfers usefully without negative interference.

free parameters (1)
  • biophysical model parameters
    These are dynamically predicted by the neural network and calibrated from data during training.
axioms (1)
  • domain assumption Biophysical crop models can be formulated differentiably for gradient-based joint optimization with a neural network.
    Required to allow the neural network to predict and optimize model parameters end-to-end.

pith-pipeline@v0.9.0 · 5706 in / 1241 out tokens · 36414 ms · 2026-05-21T10:33:26.483423+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

48 extracted references · 48 canonical work pages · 2 internal anchors

  1. [1]

    Dynamics- Based Feature Augmentation of Graph Neural Networks for Variant Emergence Prediction

    [Aawaret al., 2025 ] Majd Al Aawar, Srikar Mutnuri, Man- sooreh Montazerin, and Ajitesh Srivastava. Dynamics- Based Feature Augmentation of Graph Neural Networks for Variant Emergence Prediction. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 27793–27801, April

    work page 2025

  2. [2]

    Geoffrey Allen

    [Allen, 1994] P. Geoffrey Allen. Economic forecasting in agriculture.International Journal of Forecasting, 10(1):81–135, June

    work page 1994

  3. [3]

    Badeck, Alberte Bondeau, Kristin B ¨ottcher, Daniel Doktor, Wolfgang Lucht, J ¨org Schaber, and Stephen Sitch

    [Badecket al., 2004 ] Franz-W. Badeck, Alberte Bondeau, Kristin B ¨ottcher, Daniel Doktor, Wolfgang Lucht, J ¨org Schaber, and Stephen Sitch. Responses of spring phenol- ogy to climate change.New Phytologist, 162(2):295–309,

    work page 2004

  4. [4]

    Physics- Informed Neural Networks for Heat Transfer Problems

    [Caiet al., 2021 ] Shengze Cai, Zhicheng Wang, Sifan Wang, Paris Perdikaris, and George Em Karniadakis. Physics- Informed Neural Networks for Heat Transfer Problems. Journal of Heat Transfer, 143(060801), April

    work page 2021

  5. [5]

    Multitask Learning.Ma- chine Learning, 28(1):41–75, July

    [Caruana, 1997] Rich Caruana. Multitask Learning.Ma- chine Learning, 28(1):41–75, July

    work page 1997

  6. [6]

    Multi-Task Learning for Sequence Tagging: An Empirical Study

    [Changpinyoet al., 2018 ] Soravit Changpinyo, Hexiang Hu, and Fei Sha. Multi-Task Learning for Sequence Tagging: An Empirical Study. InInternational Conference on Com- putational Linguistics, August

    work page 2018

  7. [7]

    Empirical Evalu- ation of Gated Recurrent Neural Networks on Sequence Modeling

    [Chunget al., 2014 ] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical Evalu- ation of Gated Recurrent Neural Networks on Sequence Modeling. InNIPS 2014 Workshop on Deep Learning and Representation Learning, December

    work page 2014

  8. [8]

    [de Mattos Netoet al., 2022 ] Paulo S. G. de Mattos Neto, George D. C. Cavalcanti, Domingos S. de O. Santos J´unior, and Eraylson G. Silva. Hybrid systems using resid- ual modeling for sea surface temperature forecasting.Sci- entific Reports, 12(1):487, January

    work page 2022

  9. [9]

    [Fenget al., 2022 ] Dapeng Feng, Jiangtao Liu, Kathryn Lawson, and Chaopeng Shen. Differentiable, Learn- able, Regionalized Process-Based Models With Multi- physical Outputs can Approach State-Of-The-Art Hydro- logic Prediction Accuracy.Water Resources Research, 58(10):e2022WR032404,

    work page 2022

  10. [10]

    Ferguson, Julie M

    [Fergusonet al., 2011 ] John C. Ferguson, Julie M. Tarara, Lynn J. Mills, Gary G. Grove, and Markus Keller. Dy- namic thermal time model of cold hardiness for dormant grapevine buds.Annals of Botany, 107(3):389, January

    work page 2011

  11. [11]

    Ferguson, Michelle M

    [Fergusonet al., 2014 ] John C. Ferguson, Michelle M. Moyer, Lynn J. Mills, Gerrit Hoogenboom, and Markus Keller. Modeling Dormant Bud Cold Hardiness and Bud- break in Twenty-Three Vitis Genotypes Reveals Variation by Region of Origin.American Journal of Enology and Viticulture, 65(1):59–71, March

    work page 2014

  12. [12]

    Greer, Jens N

    [Greeret al., 2006 ] Dennis H. Greer, Jens N. W ¨unsche, Cara L. Norling, and Harry N. Wiggins. Root-zone tem- peratures affect phenology of bud break, flower cluster de- velopment, shoot extension growth and gas exchange of ’Braeburn’ (Malus domestica) apple trees.Tree Physiol- ogy, 26(1):105–111, January

    work page 2006

  13. [13]

    Phenological re- sponse to climatic change depends on spring warming ve- locity.Communications Earth & Environment, 5(1):634, October

    [Guralnicket al., 2024 ] Robert Guralnick, Theresa Crim- mins, Erin Grady, and Lindsay Campbell. Phenological re- sponse to climatic change depends on spring warming ve- locity.Communications Earth & Environment, 5(1):634, October

    work page 2024

  14. [14]

    Prediction Under Uncertainty with Error-Encoding Networks

    [Henaffet al., 2017 ] Mikael Henaff, Junbo Zhao, and Yann LeCun. Prediction Under Uncertainty with Error- Encoding Networks.arXiv:1711.04994, November

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  15. [15]

    Read, Jacob A

    [Jiaet al., 2021 ] Xiaowei Jia, Jared Willard, Anuj Karpatne, Jordan S. Read, Jacob A. Zwart, Michael Steinbach, and Vipin Kumar. Physics-Guided Machine Learning for Scientific Discovery: An Application in Simulating Lake Temperature Profiles.ACM/IMS Trans. Data Sci., 2(3):20:1–20:26, May

    work page 2021

  16. [16]

    Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang

    [Karniadakiset al., 2021 ] George Em Karniadakis, Ioan- nis G. Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Na- ture Reviews Physics, 3(6):422–440, June

    work page 2021

  17. [17]

    Theory-guided Data Science: A New Paradigm for Scientific Discovery from Data

    [Karpatneet al., 2017 ] Anuj Karpatne, Gowtham Atluri, James Faghmous, Michael Steinbach, Arindam Baner- jee, Auroop Ganguly, Shashi Shekhar, Nagiza Sama- tova, and Vipin Kumar. Theory-guided Data Science: A New Paradigm for Scientific Discovery from Data. IEEE Transactions on Knowledge and Data Engineering, 29(10):2318–2331, October

    work page 2017

  18. [18]

    [Keller and Tarara, 2010] Markus Keller and Julie M. Tarara. Warm spring temperatures induce persistent season-long changes in shoot development in grapevines.Annals of Botany, 106(1):131–141, July

    work page 2010

  19. [19]

    Smithyman, William R

    [Kelleret al., 2016 ] Markus Keller, Pascual Romero, He- mant Gohil, Russell P. Smithyman, William R. Riley, L. Federico Casassa, and James F. Harbertson. Deficit Ir- rigation Alters Grapevine Growth, Physiology, and Fruit Microclimate.American Journal of Enology and Viticul- ture, 67(4):426–435, October

    work page 2016

  20. [20]

    Knowling, Bree Bennett, Bertram Ostendorf, Seth Westra, Rob R

    [Knowlinget al., 2021 ] Matthew J. Knowling, Bree Bennett, Bertram Ostendorf, Seth Westra, Rob R. Walker, Anne Pellegrino, Everard J. Edwards, Cassandra Collins, Vinay Pagay, and Dylan Grigg. Bridging the gap between data and decisions: A review of process-based models for viti- culture.Agricultural Systems, 193:103209, October

    work page 2021

  21. [21]

    Lorenz, K.w

    [Lorenzet al., 1995 ] D.h. Lorenz, K.w. Eichhorn, H. Blei- holder, R. Klose, U. Meier, and E. Weber. Growth Stages of the Grapevine: Phenological growth stages of the grapevine (Vitis vinifera L. ssp. vinifera)—Codes and descriptions according to the extended BBCH scale.Aus- tralian Journal of Grape and Wine Research, 1(2):100– 103,

    work page 1995

  22. [22]

    Analyzing the impact of forecast errors in the planning of wine grape harvesting operations using a multi-stage stochastic model approach.arXiv:2405.19997, May

    [Milani and Cawley, 2024] Alejandro Milani and Alejan- dro Mac Cawley. Analyzing the impact of forecast errors in the planning of wine grape harvesting operations using a multi-stage stochastic model approach.arXiv:2405.19997, May

    work page arXiv 2024

  23. [23]

    Nasapower database

    [NASA, 2025] NASA. Nasapower database. https://power. larc.nasa.gov/,

    work page 2025

  24. [24]

    [Parkeret al., 2013 ] Amber Parker, Inaki Garcia De Cort´azar-Atauri, Isabelle Chuine, G ´erard Barbeau, Ben- jamin Bois, Jean-Michel Boursiquot, Jean-Yves Cahurel, Marion Claverie, Thierry Dufourcq, Laurence G ´eny, Guy Guimberteau, Rainer W. Hofmann, Olivier Jacquet, Thierry Lacombe, Christine Monamy, Hernan Ojeda, Lau- rent Panigai, Jean-Christophe Pay...

    work page 2013

  25. [25]

    Automatic differentiation in pytorch

    [Paszkeet al., 2017 ] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary De- Vito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NIPS-Workshop,

    work page 2017

  26. [26]

    Raissi, P

    [Raissiet al., 2019 ] M. Raissi, P. Perdikaris, and G. E. Kar- niadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse prob- lems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, Febru- ary

    work page 2019

  27. [27]

    Reynolds

    [Reynolds, 2022] Andrew G. Reynolds. 11 - Viticultural and vineyard management practices and their effects on grape and wine quality. InManaging Wine Quality (Second Edition), pages 443–539. Woodhead Publishing Series in Food Science, Technology and Nutrition, January

    work page 2022

  28. [28]

    Rogiers, Dennis H

    [Rogierset al., 2022 ] Suzy Y . Rogiers, Dennis H. Greer, Yin Liu, Tintu Baby, and Zeyu Xiao. Impact of climate change on grape berry ripening: An assessment of adaptation strategies for the Australian vineyard.Frontiers in Plant Science, 13:1094633, December

    work page 2022

  29. [29]

    An Overview of Multi-Task Learning in Deep Neural Networks

    [Ruder, 2017] Sebastian Ruder. An Overview of Multi-Task Learning in Deep Neural Networks.arXiv:1706.05098, June

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  30. [30]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nature Machine Intelligence, 1(5):206–215, May

    [Rudin, 2019] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nature Machine Intelligence, 1(5):206–215, May

    work page 2019

  31. [31]

    Salazar-Guti´errez and B

    [Salazar-Guti´errez and Chaves-Cordoba, 2020] M.R. Salazar-Guti´errez and B. Chaves-Cordoba. Model- ing approach for cold hardiness estimation on cherries. Agricultural and Forest Meteorology, 287:107946, June

    work page 2020

  32. [32]

    Multi-Task Learning for Budbreak Prediction

    [Saxenaet al., 2023a ] Aseem Saxena, Paola Pesantez- Cabrera, Rohan Ballapragada, Markus Keller, and Alan Fern. Multi-Task Learning for Budbreak Prediction. arXiv:2301.01815, January

    work page arXiv

  33. [33]

    [Seidelet al., 2018 ] S. J. Seidel, T. Palosuo, P. Thorburn, and D. Wallach. Towards improved calibration of crop models – Where are we now and where should we go?European Journal of Agronomy, 94:25–35, March

    work page 2018

  34. [34]

    Differentiable modeling to unify machine learning and physical models and advance Geosciences.arXiv:2301.04027, May

    [Shenet al., 2023 ] Chaopeng Shen, Alison Appling, Pierre Gentine, Toshiyuki Bandai, Hoshin Gupta, Alexandre Tartakovsky, Marco Baity-Jesi, Fabrizio Fenicia, Daniel Kifer, Xiaofeng Liu, Li Li, Dapeng Feng, Wei Ren, Yi Zheng, Ciaran Harman, Martyn Clark, Matthew Far- thing, and Praveen Kumar. Differentiable modeling to unify machine learning and physical m...

    work page arXiv 2023

  35. [35]

    Wofostgym: A crop simulator for learning annual and perennial crop management strategies

    [Solowet al., 2025 ] William Solow, Sandhya Saisubrama- nian, and Alan Fern. Wofostgym: A crop simulator for learning annual and perennial crop management strategies. InReinforcement Learning Conference (RLC) 2025,

    work page 2025

  36. [36]

    Axiomatic attribution for deep net- works

    [Sundararajanet al., 2017 ] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep net- works. InProceedings of the 34th International Confer- ence on Machine Learning (ICML 2017), pages 3319–

    work page 2017

  37. [37]

    Learning to Calibrate Battery Models in Real-Time with Deep Reinforcement Learning.Energies, 14(5):1361, January

    [Unagaret al., 2021 ] Ajaykumar Unagar, Yuan Tian, Manuel Arias Chao, and Olga Fink. Learning to Calibrate Battery Models in Real-Time with Deep Reinforcement Learning.Energies, 14(5):1361, January

    work page 2021

  38. [38]

    Athanasiadis

    [Van Breeet al., 2025 ] Ron Van Bree, Diego Marcos, and Ioannis N. Athanasiadis. Hybrid Phenology Modeling for Predicting Temperature Effects on Tree Dormancy. InPro- ceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 28458–28466, April

    work page 2025

  39. [39]

    van Diepen, J

    [van Diepenet al., 1989 ] C.a. van Diepen, J. Wolf, H. van Keulen, and C. Rappoldt. WOFOST: A simulation model of crop production.Soil Use and Management, 5(1):16– 24,

    work page 1989

  40. [40]

    Wind Power Forecasting using LSTMs

    [Vijayshankaret al., 2021 ] Sanjana Vijayshankar, Jennifer King, and Peter Seiler. Wind Power Forecasting using LSTMs. In2021 60th IEEE Conference on Decision and Control (CDC), pages 3658–3663, December

    work page 2021

  41. [41]

    Hybrid mechanistic and neural network modeling of nuclear reactors.Energy, 282:128931, November

    [Wanget al., 2023 ] Pengfei Wang, Ze Zhu, Wenlong Liang, Longtao Liao, and Jiashuang Wan. Hybrid mechanistic and neural network modeling of nuclear reactors.Energy, 282:128931, November

    work page 2023

  42. [42]

    Integrating Sci- entific Knowledge with Machine Learning for Engineering and Environmental Systems.ACM Computing Surveys, 55(4):1–37, April

    [Willardet al., 2023 ] Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar. Integrating Sci- entific Knowledge with Machine Learning for Engineering and Environmental Systems.ACM Computing Surveys, 55(4):1–37, April

    work page 2023

  43. [43]

    Agweathernet

    [WSU, 2025] WSU. Agweathernet. https://weather.wsu.edu,

    work page 2025

  44. [44]

    Predicting Key Phenological Stages for 17 Grapevine Cultivars (Vitis vinifera L.).American Journal of Enology and Viticulture, 68(1):60–72, January

    [Zapataet al., 2017 ] Diana Zapata, Melba Salazar- Gutierrez, Bernardo Chaves, Markus Keller, and Gerrit Hoogenboom. Predicting Key Phenological Stages for 17 Grapevine Cultivars (Vitis vinifera L.).American Journal of Enology and Viticulture, 68(1):60–72, January

    work page 2017

  45. [45]

    An overview of multi-task learning.National Science Review, 5(1):30–43, January

    [Zhang and Yang, 2018] Yu Zhang and Qiang Yang. An overview of multi-task learning.National Science Review, 5(1):30–43, January

    work page 2018

  46. [46]

    Appendix A: Model Architectures In this section we provide additional information on the DMC-MTL model and the variants that we used as experi- ment baselines. Appendix A.1: DMC Models The DMC-MTL model is composed of three parts, the RNN- Backbone (Figure 2a), the multi-task embedding (Figure 2b) and the interaction between the deep learning model and th...

    work page 2014

  47. [47]

    Ac- curate forecast of staple crop yields are critical to financial planning

    Appendix B.4: WOFOST Wheat Model Description As an additional testing domain to complement phenology and cold-hardiness, we consider the wheat yield domain. Ac- curate forecast of staple crop yields are critical to financial planning. The WOFOST crop growth model [van Diepenet al., 1989] is widely used to predict field level yield for many crops, includin...

    work page 1989

  48. [48]

    We choose 0.01 as the largest fac- tor as it roughly represents the largest observed daily change in the base DMC-MTL models

    We let the model parame- ters vary by a factor of 0.01, 0.001, 0.0001, and 0.00001 of the original parameter ranges. We choose 0.01 as the largest fac- tor as it roughly represents the largest observed daily change in the base DMC-MTL models. Our results in Figure 11 demonstrate that increased expres- sivity (model parameter delta scale increasing from 0....

    work page 2020