NeuralCrop.jl: A Global Crop Model Integrating Physical Mechanisms and Machine Learning

NeuralCrop.jl is an open-source project based on the Julia language. It combines traditional physical models of crop growth with neural networks to build a differentiable global gridded crop model, providing a new tool for precision agriculture and climate change research. This project aims to combine the interpretability of physical models with the fitting ability of machine learning, addressing the problems of traditional models in complex data processing and the lack of interpretability in pure ML models.

Under global climate change, accurate prediction of crop yields is crucial for food security. Traditional physical models can explain growth mechanisms but have limited ability to handle complex nonlinear relationships and massive data; pure data-driven ML methods have strong predictive power but lack interpretability and struggle to capture physical constraints. Thus, NeuralCrop.jl was born, integrating the advantages of both—retaining the physical foundation while using neural networks to handle spatiotemporal patterns.

Differentiable Programming Paradigm

Using the Julia language, leveraging the differentiable capabilities of frameworks like Flux.jl: physical parameters can be automatically optimized and calibrated, neural networks are seamlessly integrated with physical equations, supporting end-to-end gradient descent training.

Global Gridded Modeling

Unlike site-scale models, it uses a global gridded architecture: spatial resolution of 0.5°-2°, flexible time steps (daily/monthly), leveraging Julia's multi-threading and distributed computing to process large-scale data, enabling simulation of growth conditions at tens of thousands of grid points worldwide.

Introduction of Physical Constraints

Retains key physical processes: photosynthesis based on the Farquhar model, transpiration via the Penman-Monteith equation, phenological development using accumulated temperature models, carbon-nitrogen balance allocation and redistribution—ensuring results comply with biophysical laws.

Neural Network Supplements

Responsible for capturing relationships that are difficult to explicitly express in physical models: mapping from environmental conditions to crop parameters (replacing lookup tables), correction of systematic biases in physical models, and extraction of regional growth patterns from satellite remote sensing data.

Climate Change Impact Assessment

Can assess global crop yield changes under different climate scenarios; due to its physical foundation, predictions are more applicable to unobserved climate conditions.

Agricultural Policy Simulation

Policymakers can evaluate the impact of irrigation investments, the effects of adjusting planting dates, and the impact of extreme weather on food markets.

Parameter Optimization

Automatically learns optimal parameters from observational data, quickly adapts to new varieties/regions, and quantifies the impact of parameter uncertainty on predictions.

Deep Integration with the Julia Ecosystem

Integrates the SciML ecosystem (DifferentialEquations.jl) to solve differential equations, achieves GPU acceleration via CUDA.jl, and uses Zygote.jl for efficient backpropagation.

Modular Design

Physical process components can be independently replaced or upgraded, neural network structures are flexibly configurable, supporting expansion to multiple crop types and growth stages.

Current Challenges

Global gridded operation requires large amounts of high-quality input data; high-resolution simulation has high computational costs; global-scale validation lacks sufficient observational data.

Future Directions

Integrate more crop types and cropping systems; introduce reinforcement learning to optimize agricultural decisions; couple with Earth system models to achieve climate-crop bidirectional feedback; develop user-friendly visualization interfaces.

NeuralCrop.jl represents an important exploration direction in the field of crop modeling, leveraging machine learning capabilities while maintaining physical interpretability. This hybrid modeling approach is not only applicable to agriculture but also provides reference for other Earth system science problems. As the Julia ecosystem matures and computing power improves, similar physics-data fusion models are expected to play a role in more scientific fields.