SciMLSensitivity.jl：科学机器学习中的自动微分与伴随敏感性分析

SciMLSensitivity.jl is a core component of the Julia-based SciML ecosystem, providing efficient automatic differentiation and adjoint sensitivity analysis capabilities for differential equation solvers. It addresses the critical need for gradient computation in training hybrid models combining physical laws (via differential equations) and data-driven neural networks (like Neural ODEs and Universal Differential Equations). Key strengths include support for diverse equation types, modular design, and high performance.

In SciML, combining physical differential equations with neural networks (e.g., Neural ODEs, Universal Differential Equations) requires efficient calculation of loss gradients relative to model parameters. SciMLSensitivity.jl solves this by offering a modular framework for sensitivity analysis, supporting everything from simple ODEs to complex stochastic delay differential equations.

SciMLSensitivity.jl implements two main sensitivity paradigms:

1. Optimize-then-Discretize: Defines adjoint equations at the continuous level (e.g., Backsolve Adjoint, Interpolating Adjoint) with O(1) memory complexity.
2. Discretize-then-Optimize: Solves sensitivity equations on discrete time steps (generalizes ANODE methods).

It integrates seamlessly with Julia's auto-diff libraries (Zygote.jl, ReverseDiff.jl, Tracker.jl, ForwardDiff.jl). ForwardDiff uses Dual numbers directly for efficient gradient computation in some cases.

SciMLSensitivity.jl supports a wide range of equation types:

• Ordinary Differential Equations (ODEs, split ODEs)
• Stochastic Differential Equations (SDEs, SDAEs)
• Delay Differential Equations (DDEs, neutral/ algebraic variants)
• Differential Algebraic Equations (DAEs)
• Stochastic Delay Differential Equations (SDDEs, experimental)
• Mixed discrete-continuous equations (via callbacks)
• Partial Differential Equations (PDEs, via finite difference/volume or PINNs)

This makes it one of the most comprehensive sensitivity tools in SciML.

SciMLSensitivity.jl is key for training neural differential equations:

• Neural ODEs: Uses adjoint methods to avoid storing intermediate activations (O(1) memory).
• Neural SDEs: Uses Virtual Brownian Trees for O(1) memory in backpropagation.

Universal Differential Equations (UDEs) applications:

• Replace unknown physical terms with neural networks.
• Parameter estimation and model calibration.
• Automated model discovery.
• Nonlinear optimal control.

Performance: SciMLSensitivity.jl is optimized for speed—official benchmarks show it’s often orders of magnitude faster than libraries like torchdiffeq. It supports GPU acceleration (recompiles solvers to GPU, eliminates CPU-GPU data transfer).

Modularity: Users can swap ODE solvers (Sundials.jl, OrdinaryDiffEq.jl, etc.) and neural frameworks (Lux.jl, TensorFlow.jl, Torch.jl). It integrates with Bayesian packages (Turing.jl, Gen.jl) for Bayesian Neural ODEs without code changes.

SciMLSensitivity.jl is used across domains:

• Physics-informed ML: Incorporate physical constraints into neural network training.
• Drug动力学: Model drug distribution/metabolism via differential equations.
• Climate modeling: Parameter estimation and uncertainty quantification for Earth system models.
• Financial engineering: Derivatives pricing and risk modeling using SDEs.

SciMLSensitivity.jl is a leading tool for sensitivity analysis in SciML, offering a unified interface for diverse equations, efficient adjoint methods, and strong composability with Julia’s ecosystem. It empowers researchers and engineers to build and train complex hybrid models. As SciML evolves, it will play an increasingly important role in bridging physical modeling and data-driven approaches.