Equations relating a function to its derivatives, used to model how quantities change with respect to one or more independent variables; foundational to physics, engineering, biology, and modern machine-learning architectures such as neural ODEs.
Semantic Classification
Content
Compositional Relationships (Components)
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:hasPart math:OrdinaryDifferentialEquation))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:hasPart math:PartialDifferentialEquation))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:hasPart math:StochasticDifferentialEquation))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:hasPart math:BoundaryValueProblem))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:hasPart math:InitialValueProblem))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:hasPart math:NeuralODE))
Dependency Relationships
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:requires math:Calculus))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:requires math:LinearAlgebra))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:requires math:RealAnalysis))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:dependsOn math:FunctionalAnalysis))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:requires math:NumericalMethods))
Capability Relationships
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:enables math:DynamicalSystemsTheory))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:enables math:NeuralODE))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:enables math:PhysicsSimulation))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:enables math:ControlTheory))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:enables math:PhysicsInformedNeuralNetwork))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:enables math:ContinuousNormalizingFlow))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:enables math:DiffusionModel))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:supports math:ReinforcementLearning))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:supports math:Robotics))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:supports math:ClimateModelling))
Implementation Relationships
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:implements math:NavierStokesEquations))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:implements math:SchroedingerEquation))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:implements math:HeatEquation))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:uses math:FiniteElementMethod))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:uses math:RungeKuttaMethod))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:uses math:FourierAnalysis))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:uses math:LaplaceTransform))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:uses math:AutomaticDifferentiation))
Reduction Relationships
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:reducesTo math:AlgebraicEquations))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:reducesTo math:LinearSystem))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:reducesTo math:OptimisationProblem))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:contrastsWith math:AlgebraicEquations))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:relatedTo math:ProbabilityTheory))
SubClassOf(math:DifferentialEquations
ObjectSomeValuesFrom(math:relatedTo math:FourierAnalysis))
About
Differential equations constitute one of the oldest and most practically important branches of mathematics, tracing their origins to Isaac Newton and Gottfried Wilhelm Leibniz who developed calculus precisely to formalise the mathematics of change. Newton’s second law F = ma is a second-order ODE relating acceleration (the second derivative of position) to applied force; his law of gravitation defines a force field whose flow lines are trajectories satisfying this ODE. The 300-year development of the field — from Euler’s first numerical integration method (1768) through Fourier’s heat equation (1822), Cauchy-Kovalevskaya existence theory (1842–1875), Poincaré’s qualitative dynamical systems analysis (1890s), and Sobolev spaces for weak PDE solutions (1930s) to modern numerical analysis and computational PDE solvers — has made differential equations the universal language of quantitative science and engineering.
The classical taxonomy of differential equations divides the field by structural properties that determine which solution techniques are applicable:
By number of independent variables: An Ordinary Differential Equation (ODE) involves derivatives of a function with respect to a single independent variable (typically time t). A Partial Differential Equation (PDE) involves partial derivatives with respect to two or more independent variables (e.g., space x, y, z and time t). ODEs describe point-particle dynamics, chemical reaction kinetics, population ecology, and circuit theory; PDEs describe field phenomena — heat conduction, wave propagation, electromagnetism, fluid flow, and quantum mechanics.
By linearity: A linear differential equation has the form L[y] = f where L is a linear operator (no products or nonlinear functions of y or its derivatives); its solutions form a vector space and superposition applies. Nonlinear differential equations — where L contains products, powers, or nonlinear functions of y — typically cannot be solved by superposition and may exhibit qualitatively rich behaviour including multiple equilibria, periodic orbits, bifurcations, and deterministic chaos. The vast majority of practically important differential equations are nonlinear; linear equations are studied intensively because they admit complete analytical solutions and because nonlinear equations are often linearised locally (Jacobian analysis) to understand local stability.
By order: The order of a differential equation is the highest derivative that appears. First-order ODEs (y’ = f(x,y)) describe exponential growth/decay, logistic growth, and first-order filter dynamics. Second-order ODEs (y” = f(x,y,y’)) describe oscillators, waves, and gravitational orbits. Higher-order equations arise in beam mechanics (fourth-order Euler-Bernoulli beam equation) and thin-plate elasticity.
By stochasticity: Deterministic differential equations have completely specified right-hand sides. Stochastic Differential Equations (SDEs) augment the deterministic drift term with a noise term driven by Brownian motion (Wiener process) W_t: dX_t = μ(X_t, t) dt + σ(X_t, t) dW_t. The Itô calculus provides the mathematical framework for rigorous treatment of stochastic integrals and the associated chain rule (Itô’s lemma). SDEs are foundational in mathematical finance (Black-Scholes option pricing), statistical physics (Langevin dynamics), stochastic control, and — through their connection to score-based diffusion — modern Diffusion Model generative AI systems.
Most differential equations of engineering and scientific interest lack closed-form solutions. The central practical tool is Numerical Methods: algorithms that discretise a continuous differential equation into an algebraic system solvable by computer. For ODEs, the primary numerical integrators are Euler’s method (first-order explicit), the classical four-stage Runge-Kutta method (RK4, fourth-order), and adaptive Runge-Kutta-Fehlberg schemes (RK45, RK78) that automatically adjust step size to control local truncation error. For stiff ODEs — where some components decay on much faster time scales than others, requiring prohibitively small explicit time steps — implicit methods such as backward Euler, Crank-Nicolson, and SDIRK (singly diagonally implicit Runge-Kutta) are preferred. For PDEs, the principal discretisation strategies are finite differences (replace derivatives with difference quotients on a structured grid), Finite Element Method (represent the solution as a sum of piecewise basis functions, minimise the PDE residual in a weighted sense), and finite volumes (discretise integral forms of conservation laws on a mesh of control volumes). Spectral methods — expanding solutions in global basis functions (Fourier series, Chebyshev polynomials) — achieve exponential convergence for smooth solutions and are used in high-accuracy climate simulation and fluid dynamics.
Major Families
Ordinary Differential Equations (ODEs)
ODEs govern systems that evolve in time according to a state equation dx/dt = f(x, t), where x ∈ ℝⁿ is the state vector. The qualitative theory — Dynamical Systems Theory — analyses long-term behaviour without solving the equation explicitly. Key concepts include:
Fixed points (equilibria): States x* where f(x*, t) = 0 for all t. Stability of fixed points is characterised by the eigenvalues of the Jacobian ∂f/∂x at x*: negative real parts imply asymptotic stability; positive real parts imply instability.
Limit cycles: Isolated periodic orbits in the phase space. The van der Pol oscillator and Hodgkin-Huxley neuron models exhibit limit cycle dynamics.
Chaos: Sensitive dependence on initial conditions in deterministic nonlinear ODEs (Lorenz system, Rössler attractor). The Lorenz equations dx/dt = σ(y−x), dy/dt = x(ρ−z)−y, dz/dt = xy−βz with σ=10, ρ=28, β=8/3 exhibit the “butterfly attractor” — a bounded, aperiodic, non-periodic trajectory — representing deterministic chaos.
Bifurcations: Qualitative changes in system behaviour as a parameter crosses a critical value. Saddle-node, pitchfork, Hopf, and period-doubling bifurcations organise the landscape of nonlinear ODE dynamics. Bifurcation theory underpins stability analysis in Control Theory and Reinforcement Learning environment modelling.
Partial Differential Equations (PDEs)
PDEs describe continuous fields. The classical PDE triumvirate consists of:
Elliptic PDEs (e.g., Laplace equation ∇²u = 0, Poisson equation ∇²u = f): Govern steady-state equilibria. Arise in electrostatics, steady heat conduction, and fluid statics. Solutions are smooth; numerical solution by Finite Element Method or spectral methods.
Parabolic PDEs (e.g., heat equation ∂u/∂t = α∇²u): Govern diffusion and time evolution toward equilibrium. The fundamental solution is the Gaussian; parabolic PDEs are central to option pricing (Black-Scholes), image processing (anisotropic diffusion), and the mathematical analysis of Diffusion Model score functions (the Fokker-Planck equation governing probability density evolution under an SDE is a parabolic PDE).
Hyperbolic PDEs (e.g., wave equation ∂²u/∂t² = c²∇²u): Govern wave propagation and advection. Arise in acoustics, electromagnetics, seismics, and relativistic field theory. Solutions can develop discontinuities (shocks); numerical solution requires upwind schemes and Riemann solvers.
The Navier-Stokes equations — the PDEs governing viscous fluid motion — combine both parabolic (diffusion) and hyperbolic (advection) character: ρ(∂u/∂t + u·∇u) = −∇p + μ∇²u + f. Their mathematical analysis (in particular whether smooth solutions in three dimensions remain smooth for all time, the “global regularity” question) remains one of the seven Millennium Prize Problems (Clay Mathematics Institute, $1M prize).
Stochastic Differential Equations (SDEs)
SDEs of the form dX_t = μ(X_t, t)dt + σ(X_t, t)dW_t govern systems driven by continuous random noise. Itô integration (Itô 1944) and Stratonovich integration (Stratonovich 1966) provide two distinct but related interpretations of the stochastic integral, each appropriate to different physical contexts. The Fokker-Planck (Kolmogorov forward) equation governs the probability density p(x,t) of X_t: ∂p/∂t = −∂/∂x[μ(x,t)p] + ½∂²/∂x²[σ²(x,t)p]. This PDE connects SDEs to classical diffusion theory and provides the theoretical foundation for score-based Diffusion Models: by choosing the drift μ to anneal the data distribution toward noise (Ornstein-Uhlenbeck process: μ = −x), the Fokker-Planck equation describes how a complex data distribution dissolves into a Gaussian, and its reversal (the reverse-time SDE) generates novel samples from the data distribution given a learned score function ∇_x log p(x,t). This mathematical connection, articulated by Song et al. (2020, “Score-Based Generative Modeling through Stochastic Differential Equations”), provides the theoretical unification of denoising diffusion probabilistic models (DDPMs) with continuous normalizing flows and score matching into a single SDE-based framework.
Neural Differential Equations
The intersection of differential equations with deep learning has produced a family of architectures that embed neural networks as components within differential equation systems:
Neural ODEs (Chen et al. 2018, NeurIPS): Replace the discrete layer-by-layer computation of a Residual Network with a continuous ODE dx/dt = f_θ(x, t), where f_θ is a neural network. The output of the network at depth T is the solution x(T) of this ODE, computed by a numerical ODE solver (torchdiffeq library: dopri5, adams, euler). Gradients with respect to parameters θ are computed by solving a backward-time adjoint ODE, avoiding the memory cost of storing all intermediate states. Neural ODEs enable continuous-time dynamics modelling, naturally handle irregular time-series inputs (evaluate x(t) at any t), and provide a rigorous connection between depth and continuous transformations. Used in normalizing flows (FFJORD: Grathwohl et al. 2019), irregular time series (Latent ODE: Rubanova et al. 2019), and protein trajectory modelling (PHY-CHEMNODE, NeurIPS 2024).
Neural Controlled Differential Equations (CDEs) (Kidger et al. 2020): Extend Neural ODEs to inputs driven by continuous signals rather than fixed initial conditions: dz_t = f_θ(z_t) dX_t, where X_t is the input path. CDEs are the optimal architecture for irregularly-sampled time series with missing observations, outperforming RNNs and LSTMs on medical and financial time-series benchmarks. Neural CDEs motivated the Comprehensive Review of Neural Differential Equations for Time Series (IJCAI 2025).
Neural Stochastic Differential Equations (Neural SDEs) (Tzen and Raginsky 2019; Li et al. 2020): Parameterise both drift and diffusion coefficients of an SDE with neural networks. Related to latent diffusion models and variational inference; provide principled uncertainty quantification for sequential predictions. Stable neural SDEs (ICLR 2024) address the challenge of maintaining stochastic stability during training on irregular time-series data.
Physics-Informed Neural Networks (PINNs) (Raissi, Perdikaris, Karniadakis 2019): Train neural networks to solve PDEs by including the PDE residual as a loss term. For a PDE Lu = f(x,t), a network u_θ(x,t) is trained to minimise L_PDE = ∫|L[u_θ] − f|² + L_BC (boundary conditions) + L_IC (initial conditions). PINNs can solve forward problems (given PDE and boundary conditions, find u) and inverse problems (given observations of u, infer PDE coefficients). Particularly valuable for problems where measurement data is sparse, physics is known, and traditional discretisation is intractable (high dimensions, complex geometries). The ConFIG method (ICLR 2025) addresses gradient conflicts between competing PDE and boundary loss terms.
Flow Matching and Continuous Normalizing Flows (Lipman et al. 2022; Liu et al. 2022): Frame generative modelling as learning a velocity field v_θ(x,t) defining an ODE dx/dt = v_θ(x,t) that transports a source distribution (Gaussian) to a target distribution (data). The ODE is solved with a neural network velocity field, trained by matching the velocity field to an interpolated target trajectory (flow matching). This framework unifies continuous normalizing flows, score-based diffusion, and rectified flows; Flow Matching achieves state-of-the-art performance in image, audio, protein, and robotic trajectory generation (2024–2025).
Use Cases
Computational Fluid Dynamics (CFD) and Engineering
Numerical solution of the Navier-Stokes PDEs is the computational core of aerodynamic design (aircraft wings, wind turbines, racing cars), HVAC optimisation, turbomachinery analysis, and offshore oil platform design. Finite Element Method and finite-volume codes (OpenFOAM, ANSYS Fluent, Star-CCM+) solve discretised Navier-Stokes on unstructured meshes with 10⁶–10⁹ unknowns, requiring HPC clusters. Physics-Informed Neural Networks offer surrogate models that reduce evaluation time from hours to milliseconds after offline training, enabling real-time optimisation in design loops. The University of Southampton has an active research programme on PINNs as environmentally friendly alternatives to traditional CFD, achieving CFD-level accuracy at 100–1,000,000× lower computational cost for specific problem classes.
Climate and Earth System Modelling
Global atmospheric circulation models discretise the primitive equations (a PDE system derived from Navier-Stokes with Coriolis terms, thermodynamics, and moisture physics) on a spherical grid at 10–100 km horizontal resolution. The UK Meteorological Office Unified Model (UM) is a global PDE solver running on 140,000+ CPU cores at the Exeter Cray XC40 supercomputer. Machine learning is increasingly integrated: Neural ODEs and PINNs provide data-driven closure terms for subgrid-scale turbulence parameterisation that is too costly to resolve explicitly; graph neural network weather models (GraphCast, Pangu-Weather) learn implicit ODEs over graph-structured meteorological fields, achieving 10-day forecast accuracy competitive with traditional NWP at 1,000× lower inference cost.
Medical and Biological Modelling
ODEs and PDEs govern pharmacokinetic/pharmacodynamic (PK/PD) models describing drug absorption, distribution, metabolism, and excretion; neuron firing dynamics (Hodgkin-Huxley equations — four coupled nonlinear ODEs for membrane voltage and ion channel gating variables); tumour growth models (logistic, Gompertz, reaction-diffusion PDEs); and epidemic compartment models (SIR, SEIR — systems of ODEs coupling susceptible, exposed, infected, recovered populations). The PHY-CHEMNODE framework (NeurIPS 2024) applies neural ODE + PINN hybrids to combustion chemistry kinetics, learning accurate ODE systems for 100+ coupled species from sparse measurement data. Neural SDEs model protein-ligand binding dynamics (Nature Communications 2025).
Robotics and Control
Control Theory for robotic systems is fundamentally differential-equation-based: the robot’s forward dynamics are an ODE relating joint torques to joint accelerations; the closed-loop control law must stabilise this ODE around desired trajectories. Model Predictive Control (MPC) — the dominant industrial control paradigm — repeatedly solves a constrained ODE optimisation problem at each time step. Reinforcement Learning environments for robotic simulation (MuJoCo, Isaac Gym, Brax) are numerical ODE integrators; learned policies interact with these ODE systems. Neural ODE-based dynamics models provide differentiable simulation for gradient-based trajectory optimisation without handcrafted analytical dynamics.
Financial Mathematics
The Black-Scholes PDE ∂V/∂t + ½σ²S²∂²V/∂S² + rS∂V/∂S − rV = 0 and its extensions (stochastic volatility models: Heston SDE; local volatility models: Dupire PDE) are the bedrock of derivatives pricing in quantitative finance. American option pricing requires solving a PDE with a free boundary (early exercise constraint), a parabolic obstacle problem. Deep learning surrogates for PDE-based option pricing (Deep Galerkin Method, Deep BSDE method) solve high-dimensional PDEs with neural networks by rewriting the PDE as a backwards stochastic differential equation (BSDE), enabling pricing of basket options and path-dependent derivatives in 100+ dimensions that conventional numerical methods cannot handle.
Generative AI and Score-Based Diffusion
The theoretical foundation of Diffusion Models — the dominant paradigm for image, audio, video, and protein structure generation as of 2025 — is the SDE framework. The forward diffusion process (adding noise to data) is an Ornstein-Uhlenbeck SDE; the reverse process (denoising to generate samples) is the reverse-time SDE whose drift depends on the score function ∇_x log p_t(x). Denoising Score Matching trains a neural network to estimate this score; the resulting reverse-time SDE or its deterministic ODE counterpart (probability flow ODE) generates samples by numerical integration. This connection to SDEs provides both a theoretical analysis of diffusion model expressiveness and practical tools: adaptive ODE/SDE solvers (DPM-Solver, DPM-Solver++) accelerate sampling by 10–100× versus naive Euler discretisation.
Academic Context
The mathematical theory of differential equations is one of the deepest and most richly developed areas of pure and applied mathematics, with major contributions spanning three centuries. Existence and uniqueness theory (Picard-Lindelöf theorem for ODEs; Cauchy-Kovalevskaya theorem for analytic PDEs; Sobolev space theory for weak PDE solutions, Ladyzhenskaya 1963; Evans PDE textbook 1998) provides the foundational guarantees that underpins all computational approaches.
Numerical analysis of differential equations has been developed by legendary figures including Runge, Kutta, Richardson, Crank, Nicolson, Lax, Richtmyer, and Gear, with modern synthesis in the textbooks of Hairer, Nørsett, and Wanner (Solving ODEs I-II, 1987–1996) and Trefethen and Bau (Numerical Linear Algebra, 1997). The modern adaptive ODE solver dopri5 (Dormand-Prince, 1980), used as the default solver in torchdiffeq, is a result of this tradition.
Stochastic calculus was developed independently by Kiyosi Itô (1944, Japan) and independently systematised by Stratonovich (1966, USSR). The Feynman-Kac theorem (Kac 1949) links SDEs to PDEs (parabolic equations), enabling probabilistic representations of PDE solutions and forming the theoretical bridge between the heat equation, Brownian motion, and Gaussian process regression.
The Neural ODE paper (Chen, Rubanova, Bettencourt, Duvenaud, NeurIPS 2018) is the landmark work initiating the machine learning intersection of the field, with 7,000+ citations by 2025. The score-based SDE paper (Song, Sohl-Dickstein, Kingma, Kumar, Ermon, Poole, ICLR 2021) provided the theoretical unification of diffusion models, receiving the ICLR Outstanding Paper Award. The Physics-Informed Neural Networks paper (Raissi, Perdikaris, Karniadakis, JCP 2019) catalysed a decade of PDE-ML research with 12,000+ citations.
Key academic groups at the intersection of differential equations and machine learning include: Patrick Kidger (Oxford → Google DeepMind, neural CDEs and SDEs); Georg Bökman (Chalmers, neural SDEs for time series); George Em Karniadakis (Brown University, PINNs and scientific ML); Siddhartha Mishra (ETH Zürich, neural operators for PDEs); David Duvenaud (Toronto, Neural ODEs); Yann LeCun/Ari Morcos (Meta FAIR, continuous-depth networks); Yang Song (Stanford → OpenAI, score-based diffusion); Chris Rackauckas (MIT, DifferentialEquations.jl Julia ecosystem, universal differential equations).
Current Landscape (2026)
Physics-Informed Neural Networks (PINNs): PINNs have evolved from proof-of-concept (2019) to production-applicable tools for specific engineering problem classes. The key advance is addressing gradient conflicts (ConFIG, ICLR 2025) and scalability to large 3D transient problems via adaptive loss balancing and domain decomposition (XPINNs). University of Southampton, Imperial College, and University of Exeter are among the leading UK groups deploying PINNs as CFD surrogates for wind farm design and tidal energy optimisation. Neural Concept (France/UK) provides commercial PINNs for aerodynamic shape optimisation, used by automotive OEMs including Renault and Airbus.
Neural ODE / Flow Matching: Flow Matching (Lipman et al. 2022; Liu et al. 2022) has supplanted score-based diffusion as the dominant paradigm for continuous generative modelling, used in Stable Diffusion 3 (Escher flow), Meta’s Voicebox and AudioBox, and protein structure prediction (ESMFold 2 ablations). Flow Matching’s connection to Neural ODEs (the velocity field defines an ODE) is explicit; the framework is theoretically unified with CNFs and score matching, achieving state-of-the-art results in image, audio, molecule, and robotic trajectory generation with 10–100× fewer function evaluations than standard diffusion.
Neural Controlled Differential Equations for Time Series: The comprehensive review of neural differential equations for time series (IJCAI 2025) confirmed Neural CDEs as the state-of-the-art approach for medical irregularly-sampled data (MIMIC-III ICU predictions), financial time series (order book modelling), and physics simulation output surrogates. Stable Neural SDEs (ICLR 2024) addressed the instability of unconstrained SDE training on long sequences.
Scientific Machine Learning (SciML) Ecosystem: The Julia SciML organisation (Rackauckas et al.) provides the most mature differential equations + ML ecosystem: DifferentialEquations.jl (the world’s most feature-complete ODE/SDE/DDE/DAE solver library), DiffEqFlux.jl (Neural ODE training), ModelingToolkit.jl (symbolic-numeric ODE modelling), and NeuralPDE.jl (PINNs). This ecosystem is used by pharmaceutical companies for PK/PD modelling, national labs for plasma physics simulation, and climate centres for hybrid NWP-ML models.
Climate and Weather: The Met Office (Exeter) and ECMWF (Reading) both have active programmes integrating ML with differential equation-based NWP: Met Office AICON project (2024–2027, £8M) develops neural parameterisations of cloud microphysics (replacing empirical ODE schemes with neural ODEs trained on high-resolution LES data); ECMWF’s machine learning model AIFS (Artificial Intelligence Forecasting System, 2024) uses a graph neural network implicitly learning the finite-time map of the atmospheric ODE. Both complement rather than replace traditional PDE-based NWP at coarser scales.
UK Context
Academic Research: The University of Oxford Mathematical Institute (Patrick Kidger group, now at Google DeepMind; Arnaud Doucet group — sequential Monte Carlo and SDEs) produced foundational Neural CDE and SDE work. Oxford’s numerical analysis group (Endre Süli, Christoph Reisinger) has contributed to backward SDE solvers for high-dimensional option pricing. University of Cambridge DAMTP (Department of Applied Mathematics and Theoretical Physics) hosts the Fluid Dynamics group (Grae Worster, Stuart Dalziel) working on moving-boundary PDEs for solidification and melting (climate-relevant), and the Applied Probability group (Frank Kelly, Geoffrey Grimmett) contributing to SDE theory. Imperial College Mathematics (Andrew Stuart, formerly — now Caltech; Grigoris Pavliotis on SDEs and diffusion) has published extensively on interacting particle systems and mean-field limits of SDEs relevant to neural network training dynamics. UCL Computer Science (Brooks Paige, Marc Deisenroth) applies Neural ODEs and probabilistic differential equation models to data-efficient robotics and scientific simulation surrogate modelling. Heriot-Watt University (Edinburgh) has a strong applied mathematics programme including PDE-based image processing and inverse problems. University of Warwick Mathematics Institute (Martin Hairer — Fields Medal 2014 for singular SPDEs; Stefan Adams — statistical mechanics SDEs) is among the world’s leading groups for rigorous SDE and SPDE theory.
Industrial Applications in Northern England: BAE Systems (Warton, Lancashire) uses differential equation-based structural dynamics solvers (Nastran FEM, LS-DYNA) and is actively evaluating PINNs for combat aircraft structural fatigue prediction as a surrogate for high-fidelity FEM runs, in collaboration with the University of Sheffield INSIGNEO institute. Rolls-Royce (Derby) applies PDEs extensively for turbine aerothermal simulation (HYDRA CFD code, developed internally since 1990s) and is integrating neural PDE surrogates (UQ-PDE) under the UKRI Prosperity Partnership programme (Derby/Cambridge, 2023–2027, £12M). Siemens Energy (Lincoln) uses FEM-based PDE solvers for power transformer thermal analysis and has piloted PINN-based surrogate models for real-time transformer health monitoring. AMRC Sheffield (Advanced Manufacturing Research Centre) applies ODE-based machining dynamics models (chatter prediction — a stability boundary ODE problem) with neural ODE surrogates enabling real-time adaptive machining control in aerospace titanium machining. National Grid (Warwick) models electricity grid frequency dynamics as a large ODE system (swing equations) and uses data-driven ODE identification (SINDy — Sparse Identification of Nonlinear Dynamics) for grid anomaly detection and stability margin estimation.
Policy and Funding: The UKRI EPSRC Mathematical Sciences Programme has funded several differential equations + ML intersections: “Universal Approximation via Neural Differential Equations” (Edinburgh/Oxford, £2.3M, 2023–2026); “Scientific Machine Learning for Engineering PDEs” (Cambridge/Imperial, £3.1M, 2024–2027). The Alan Turing Institute’s programme on Mathematical Sciences and Data Science includes a working group on Neural Operators and PDEs (co-chaired by researchers from Warwick, Oxford, and Imperial). EPSRC’s Living with Environmental Change programme funds NWP-ML hybrid models for flood prediction (Edinburgh, Bristol, CEH Wallingford), combining catchment hydrology PDEs with Neural ODE data assimilation.
Future Directions (2026–2030)
Near-Term (2026–2027): Neural operators (FNO — Fourier Neural Operator; DeepONet) will mature from research tools to commercial PDE solver replacements for parametric families of PDEs, with operator-learning packages (neuraloperator Python library, OpenFOAM-ML coupling) integrated into standard CFD workflows. PINN scalability to full 3D transient Navier-Stokes problems will be achieved through domain decomposition + GPU parallelism, enabling real-time CFD at engineering design accuracy for moderate Reynolds-number flows. Neural SDE samplers will become the standard foundation for diffusion-based generative models, replacing approximate DDPM discretisations with principled SDE/ODE integration with theoretical error control.
Medium-Term (2027–2029): Foundation models for scientific simulation — large neural operators pre-trained on corpora of PDE solutions across multiple physics domains (fluid, thermal, structural, electromagnetic) — will enable few-shot adaptation to new geometries and boundary conditions, analogous to LLM foundation models for language. Universal Differential Equations (UDEs; Rackauckas et al. 2021) — hybrid models combining known mechanistic ODE terms with neural network unknowns — will become the standard modelling approach in pharmacology, ecology, and climate science, enabling interpretable ML models grounded in physical conservation laws. Quantum differential equation solvers (HHL algorithm and its variants) will demonstrate exponential speedups over classical FEM for specific PDE problem classes on fault-tolerant quantum computers (50–100 logical qubit era, 2028–2030 projection).
Long-Term (2029–2030+): Fully differentiable physics engines — coupling fluid, structural, thermal, electromagnetic, and chemical PDEs in a single differentiable simulation framework — will enable end-to-end gradient-based co-design of complex engineered systems (aircraft, battery cells, fusion reactors) from sensor observations to design parameters without intermediate human-specified models. SDE-based world models for embodied AI will learn stochastic environment dynamics from interaction data, enabling model-based Reinforcement Learning with principled uncertainty quantification for safety-critical robotics (surgical robots, autonomous vehicles). Neuromorphic computing architectures (Intel Loihi 3, IBM NorthPole successors) will implement continuous-time Neural ODE inference directly in spike-timing hardware, achieving 1000× energy efficiency over GPU-based discrete RNN inference for edge robotics and IoT sensor processing.
Research and Literature
- Raissi, M., Perdikaris, P., & Karniadakis, G. E. (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378, 686–707.
- Chen, R. T. Q., Rubanova, Y., Bettencourt, J., & Duvenaud, D. (2018). Neural Ordinary Differential Equations. NeurIPS 2018. https://arxiv.org/abs/1806.07366
- Song, Y., Sohl-Dickstein, J., Kingma, D., Kumar, A., Ermon, S., & Poole, B. (2021). Score-Based Generative Modeling through Stochastic Differential Equations. ICLR 2021. https://arxiv.org/abs/2011.13456
- Kidger, P., Morrill, J., Foster, J., & Lyons, T. (2020). Neural Controlled Differential Equations for Irregular Time Series. NeurIPS 2020. https://arxiv.org/abs/2005.08926
- Lipman, Y., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., & Le, M. (2022). Flow Matching for Generative Modeling. ICLR 2023. https://arxiv.org/abs/2210.02747
- Grathwohl, W., Chen, R. T. Q., Betterncourt, J., Sutskever, I., & Duvenaud, D. (2019). FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models. ICLR 2019. https://arxiv.org/abs/1810.01367
- Rubanova, Y., Chen, R. T. Q., & Duvenaud, D. (2019). Latent Ordinary Differential Equations for Irregularly-Sampled Time Series. NeurIPS 2019. https://arxiv.org/abs/1907.03907
- Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., & Anandkumar, A. (2021). Fourier Neural Operator for Parametric Partial Differential Equations. ICLR 2021. https://arxiv.org/abs/2010.08895
- Haier, E., Nørsett, S. P., & Wanner, G. (1987–1996). Solving Ordinary Differential Equations I–II. Springer-Verlag.
- Evans, L. C. (1998). Partial Differential Equations. American Mathematical Society (Graduate Studies in Mathematics).
- Rackauckas, C., et al. (2021). Universal Differential Equations for Scientific Machine Learning. arXiv. https://arxiv.org/abs/2001.04385
- Tzen, B., & Raginsky, M. (2019). Neural Stochastic Differential Equations: Deep Latent Gaussian Models in the Diffusion Limit. arXiv. https://arxiv.org/abs/1905.09883
- Kidger, P. (2022). On Neural Differential Equations (DPhil Thesis, Oxford). https://arxiv.org/abs/2202.02435
- Hairer, M. (2014). A theory of regularity structures. Inventiones Mathematicae, 198(2), 269–504. (Fields Medal work on singular SPDEs.)
- Dormand, J. R., & Prince, P. J. (1980). A family of embedded Runge-Kutta formulae. Journal of Computational and Applied Mathematics, 6(1), 19–26.
- Black, F., & Scholes, M. (1973). The Pricing of Options and Corporate Liabilities. Journal of Political Economy, 81(3), 637–654.
- Lorenz, E. N. (1963). Deterministic Nonperiodic Flow. Journal of the Atmospheric Sciences, 20(2), 130–141.
- Poincaré, H. (1890). Sur le problème des trois corps et les équations de la dynamique. Acta Mathematica, 13, 1–270.
- Ho, J., Jain, A., & Abbeel, P. (2020). Denoising Diffusion Probabilistic Models. NeurIPS 2020. https://arxiv.org/abs/2006.11239
- Comprehensive Review of Neural Differential Equations for Time Series Analysis. (2025). IJCAI 2025. https://www.ijcai.org/proceedings/2025/1179.pdf
- Stable Neural Stochastic Differential Equations in Analyzing Irregular Time Series Data. (2024). ICLR 2024. https://proceedings.iclr.cc/paper_files/paper/2024/file/a61023ce36d21010f1423304f8ec49af-Paper-Conference.pdf
- Physics-Informed Neural Controlled Differential Equations for Scalable Long Horizon Multi-Agent Motion Forecasting. (2025). https://arxiv.org/html/2510.00401v1
- Frontiers in AI — Implementing PINNs with Deep Learning for Differential Equations. (2026). Frontiers in Artificial Intelligence. https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2026.1717117/full
- University of Southampton. Physics Informed Neural Networks as Environmentally Friendly Alternative to CFD. https://www.southampton.ac.uk/study/postgraduate-research/projects/physics-informed-neural-networks-as-an-environmentally
- Learning PDE Solvers with Physics and Data: Unifying PINNs and Neural Operators. (2026). https://arxiv.org/pdf/2601.14517
- PHY-CHEMNODE: Combustion Kinetics via Neural ODE + PINN Hybrid. (2024). NeurIPS 2024.
- Mishra, S., & Molinaro, R. (2023). Estimates on the generalization error of physics-informed neural networks for approximating PDEs. IMA Journal of Numerical Analysis, 43(1), 1–43.