From Navier-Stokes to Neural Nets: The CFD Solver Innovation Arc

CFD solvers are numerical simulation engines that discretize and solve the governing equations of fluid mechanics — primarily the Navier-Stokes equations — to predict flow fields, heat transfer, and pressure distributions across complex geometries. The technology has become indispensable across aerospace, energy, automotive, and data centre industries as pressure mounts to reduce physical prototyping costs and accelerate design cycles. The patent record spanning 1966 to 2026 tells a story of four distinct generational shifts, each compressing the time between simulation and design decision.

The timeline begins with a 1966 GB patent from United Aircraft Corporation on jet engine performance analysis using analog computing — the earliest precursor in this dataset. The foundational phase through 2005 established core numerical methods; Mazda Corporation began filing a dense cluster of 1D/3D CFD engine performance prediction patents in 2004–2005, establishing hybrid dimensional CFD modeling as a mature sub-field. The development phase (2007–2015) saw ExxonMobil file its intelligent performance assistant simulation family across at least seven jurisdictions, representing the first systematic commercial-scale effort to automate solver parameter selection. The acceleration phase (2016–2022) brought heterogeneous hardware architectures and open-source integration environments into the patent record. The frontier phase (2023–2026) — where active and pending patents are now concentrated — is defined by ML-embedded solver workflows, AI-updated virtual wind tunnels, digital twin-driven simulation, and real-time thermal CFD for data centres.

This trajectory reflects a broader industry dynamic: according to WIPO, simulation and modelling technologies have been among the fastest-growing patent categories globally over the past decade, driven by the convergence of high-performance computing and AI. The CFD solver landscape is a microcosm of this trend, with the competitive frontier shifting from numerical accuracy to workflow automation and real-time inference.

ML-Native Solver Selection: The End of Expert Configuration

The core bottleneck in CFD solver performance is that optimal solver-preconditioner-smoother combinations for large sparse matrix systems are problem-dependent — and historically required specialist knowledge to configure correctly. Two distinct patent strategies are now attacking this bottleneck simultaneously, with significant IP implications for commercial simulation software vendors.

ExxonMobil’s approach — filed across WO, US, EP, CA, IN, NO, BR, and CN jurisdictions between 2007 and 2012 — establishes a runtime adaptive model: an intelligent performance assistant initializes the simulator, autonomously selects algorithm and parameter sets for solving the numerical matrix, and auto-adjusts when runtime performance falls below a threshold. This reactive architecture was the first systematic commercial-scale effort to remove expert configuration from the solver loop.

Tata Consultancy Services’ 2024 filings (US, EP, IN) represent a structural evolution: rather than adapting at runtime, a multi-class classification ML model is trained on CFD simulation input parameters — not matrix properties — to predict the fastest solver-preconditioner-smoother combination before the simulation runs. This eliminates the matrix characterization intermediate step entirely, shifting solver selection from reactive to proactive. These claims are currently pending across three major jurisdictions, meaning that if granted broadly, they could establish significant blocking positions on the use of ML classifiers for CFD solver configuration in commercial simulation software.

“The ML pre-selection approach shifts solver selection from reactive runtime adaptation to proactive pre-simulation ML inference — a structural change in how CFD solver pipelines will be orchestrated.”

A third approach comes from Talos Innovation ApS (Denmark), whose 2023 CA and GB filings target time-step convergence specifically: a predictive model pre-trained via statistical or machine learning methods predicts the initial guess for the solver at each subsequent time step, reducing iteration counts in transient CFD simulations. The application domain is subterranean fluid simulation — reservoir and subsurface flow — where transient solver convergence is a primary computational cost driver.

A 2025 JP filing from SUPWAT Co., Ltd. extends this ML integration further downstream: the system feeds numerical analysis model results and domain-expert knowledge into a large language model to generate interpretations and recommendations for the next analysis step. This is the first signal in this dataset of generative AI being integrated into CFD post-processing workflows — a direction that researchers tracking AI’s role in scientific simulation, as documented by Nature, will recognise as an early but consequential development.

GPU, FPGA, and Many-Core: The Hardware Acceleration Race

Hardware-level CFD innovation is bifurcating along two trajectories: heterogeneous cluster orchestration that dispatches CFD task blocks across CPU, GPU, and FPGA resources; and architecture-specific optimization that ports established solvers — particularly OpenFOAM — to novel many-core processors including China’s Sunway platform.

Hangzhou Yuansuan Technology’s 2017 CN patent describes a heterogeneous cluster system that intelligently assigns CFD task blocks to CPU, GPU, or FPGA nodes based on task attributes and hardware floating-point performance, connected via InfiniBand with a PBS resource manager. The system explicitly targets the computational bottleneck in CFD simulation — not solver algorithm design, but execution infrastructure. This distinction is critical for freedom-to-operate analysis: the open-source OpenFOAM solver itself is not protected, but the execution infrastructure on novel hardware is.

Nanjing University of Aeronautics and Astronautics’ 2024 US patent takes a different approach: GPU-native computation for moving overset mesh blocks in helicopter rotor flow field simulation. The GPU computes flow field information on moving overset mesh blocks using CFD methods, while the CPU handles interpolation relationship determination and data exchange between mesh blocks. This CPU-GPU division of labour for dynamic mesh topologies is an active frontier driven by urban air mobility and next-generation propulsion design requirements.

Two patents from Chinese universities target China’s Sunway many-core processor specifically. Qingdao Marine Science and Technology National Laboratory’s 2022 CN patent ports OpenFOAM’s SmoothSolver to the Sunway architecture by converting LDU sparse matrix format to CSR format for DMA-optimized data transfer across master and slave core groups. Xi’an Jiaotong University’s 2018 CN patent targets the Sunway TaihuLight supercomputer with block-based many-core parallelization, DMA transfer optimization, double-buffering, SIMD vectorization, and register communication optimization for fluid machinery simulation programs. Together, these filings are part of a coordinated sovereign CFD hardware-software stack — a strategic dimension that R&D teams competing in the Chinese market should map carefully, as noted by OECD research on state-directed technology investment patterns.

Surrogate and Reduced-Order Models: Run CFD Once, Generalise Everywhere

Surrogate and reduced-order model patents are architecturally convergent across domains: run a full CFD simulation at reference conditions, then use mathematical or ML-based techniques to predict outcomes for new configurations without re-executing the full solver. The delta-adaptation pattern — predicting the difference from a known baseline rather than the absolute value — is emerging as the dominant architectural approach in the 2024–2026 filing cohort.

Tata Consultancy Services’ 2015 IN patent established an early template: a fixed number of full CFD simulations at reference operating conditions establish influence mass fractions, then Proper Orthogonal Decomposition (POD) predicts temperature, mass flow rate, and thermal influence for arbitrary data centre cooling configurations without re-running the CFD solver. This approach — documented in the data centre thermal management literature published by IEEE — demonstrated that CFD-derived surrogate modeling could enter IT infrastructure planning workflows.

Dell Products’ 2025 US and CN filings represent the commercial maturation of this concept: a trained ML model determines benchmark CFD for a baseline server configuration, then adapts it to target servers with different component configurations by modeling the difference — achieving real-time thermal simulation without full re-execution of the CFD solver. General Motors’ 2025 CN filing on an AI virtual wind tunnel takes the same delta-adaptation principle to aerodynamic geometry prediction: the system automatically generates new training model data when existing training geometry data is absent, updates the model database incrementally, and integrates trained models into the CFD simulation module.

In life sciences, Amgen’s 2023 KR patent trains a surrogate ML model on a training set of CFD simulation results for a steady-state mixing tank, then uses it to predict mixing quality for new configurations without running the full CFD model. Regeneron Pharmaceuticals (JP, 2024; CN, 2023) uses CFD simulation results — including steady-state and transient flow analysis — as evaluation criteria for developing predictive mixing protocol models for bioreactor optimization. The architectural commonality across these filings — data centre thermal (Tata, Dell), automotive aerodynamics (GM), bioreactor mixing (Amgen, Regeneron), and reservoir flow (ExxonMobil) — creates both freedom-to-operate risks and licensing opportunities that cross sector boundaries.

Domain Applications: Aerospace, Automotive, Data Centres, and Beyond

Aerospace represents the highest-activity application domain by recent filing count in this dataset. China’s Institute of High Speed Aerodynamics (CARDC) filed on batch CFD job construction and submission for multi-variable aerospace aerodynamic databases (CN, 2020). Dassault Systemes Americas filed on automated turbofan engine CFD simulation — including automated solver input file generation from CAD models and cloud-platform execution — in both CN (2025) and JP (2025) jurisdictions. Nanjing University of Aeronautics and Astronautics filed a GPU-based overset mesh CFD system for helicopter rotor flow fields (US, 2024) and an adaptive MR-WENO/linear upwind hybrid reconstruction method for complex aero-engine flow problems (CN, 2025). Korea Aerospace Industries filed an open-source aerodynamic CFD analysis system for aircraft exterior flow (KR, 2016).

In automotive powertrain, Mazda Corporation’s 12+ JP patent filings between 2004 and 2012 remain the most prolific single-assignee cluster in this dataset. The core innovation couples a fast 1D CFD model for full intake-exhaust flow with a selective 3D CFD model applied only to critical sections where accuracy demands justify the computational cost. More recent entries include First Automobile Works (China) filing an automated CFD modeling and analysis system based on ANSA and StarCCM+ platforms for vehicle exterior flow field simulation (CN, 2021), and Kia Motors’ engine virtual test environment using 1D physics-based models combined with data-driven models for transient engine response (CN, 2020/2023).

Data centre thermal management is among the fastest-growing application sub-domains in recent filings. Schneider Electric IT (CN, 2014) filed on potential flow-based CFD for data centre airflow prediction using unstructured mesh. Tata Consultancy Services (IN, 2015) applied POD-based reduced-order CFD surrogate for real-time thermal prediction. Dell Products filed two patents in 2025 (US and CN) on ML-adapted CFD for real-time server thermal simulation. In nuclear safety, Enesys Co., Ltd. (KR, 2010) applied 3D CFD thermal-hydraulic analysis to nuclear reactor vessel safety. In life sciences, Regeneron Pharmaceuticals (JP, 2024; CN, 2023) and Amgen (KR, 2023) both use CFD simulation results to train surrogate predictive models for bioreactor mixing optimization — a domain where CFD has historically been underutilised relative to its potential, as noted in PatSnap’s innovation intelligence resources.

Strategic Implications for IP and R&D Teams

Five strategic patterns emerge from the 2023–2026 filing cohort that IP strategists and R&D leaders should act on now, before the pending claims in this dataset are granted and competitive positions harden.

1. Monitor Tata Consultancy Services’ Pending ML Solver Claims

The ML pre-selection approach for CFD solver configuration is currently in pending status across US, EP, and IN jurisdictions. These claims, if granted broadly, could establish significant blocking positions on the use of ML classifiers for CFD solver configuration in commercial simulation software. Any vendor building ML-assisted solver orchestration should conduct a freedom-to-operate analysis against these pending claims before committing to a product architecture.

2. Map China’s Sovereign CFD Stack

The combination of Sunway-optimized OpenFOAM solvers, heterogeneous cluster CFD dispatch systems, aerospace batch submission frameworks, and domestic simulation platform patents represents a coordinated, domestically-oriented CFD capability build. R&D teams targeting the Chinese market — or competing with Chinese aerospace and automotive simulation tools — should map this filing cluster carefully using tools such as PatSnap’s competitive intelligence platform.

3. Assess Cross-Sector Surrogate Model Exposure

Tata Consultancy Services (data centre thermal, POD), Dell (server thermal, ML delta-adaptation), Amgen (bioreactor mixing, ML surrogate), and ExxonMobil (reservoir flow, intelligent assistant) all protect variants of the “run CFD once, use ML to generalise” architectural pattern. The commonality creates both freedom-to-operate risks and licensing opportunities across sectors — a cross-industry IP landscape that requires sector-spanning claim mapping rather than domain-specific analysis.

4. Recognise the Automated Workflow as the New Competitive Battleground

Patents from Dassault Systemes Americas (CAD-to-CFD automation for turbofan engines), General Motors (AI virtual wind tunnel with self-updating design of experiments), and First Automobile Works (ANSA-StarCCM+ automated workflow) collectively indicate that competitive value is shifting from solver accuracy per se to end-to-end workflow automation. Vendors and R&D organisations that control automated pre-processing, solver orchestration, and post-processing pipelines will have significant leverage over those offering only solver technology — a pattern consistent with platform dynamics observed across scientific software markets by IEEE.

5. Track LLM Integration as an Emerging Frontier

SUPWAT Co., Ltd.’s 2025 JP filing — feeding CFD results and domain-expert knowledge into a large language model to generate interpretations and recommendations — is the first signal in this dataset of generative AI integration into CFD post-processing. While a single filing is not a trend, it is an early indicator that the competitive frontier may extend beyond solver orchestration to AI-assisted interpretation and next-step recommendation. IP teams should begin monitoring this sub-space before it becomes crowded.