About
DeepUnity is an independent research effort on fast, physics-faithful models for offshore pipe problems. It is built and maintained by one engineer, and every tool on it comes from a study with its data, method and error metrics written down.
Subsea jumper design is still done case by case in finite-element software, and a single FEED-level configuration can take days of iteration. The jumper project asks how much of that a model trained on the same physics can do in seconds, without hiding the engineering checks inside a black box: the networks predict per-element results, and everything after them is the same arithmetic a verification engineer would run.
The guided-wave project takes the opposite route. Instead of a surrogate, it makes the finite-element solver itself fast and differentiable on a GPU, so that thousands of pipe geometries and pressures can be swept at once and inverse problems can be solved with exact gradients.
Chi Yang, Ph.D.
Engineer working across offshore structures, machine learning and high-performance computing
Seven-plus years bridging machine learning, data science and engineering, with production AI work in the offshore and transportation industries.
- Offshore structures
- Subsea systems and pipelines, strength and fatigue assessment, nonlinear finite-element analysis
- Machine learning
- Deep learning, surrogate and physics-informed modelling, anomaly detection
- Scientific computing
- High-performance numerical methods, GPU computing, optimization and inverse problems
- Engineering software
- Digital twins, data platforms, production machine-learning systems, AI agents
Timeline
-
DeepUnity v2
This site: two projects, CPU-only tools, design search back online.
-
GPU-SAFE
Differentiable GPU-batched SAFE solver for guided waves in pressurized pipes.
-
Transformer fatigue surrogate
Sectional forces per element, DNV-RP-C203 damage, paired operating-state validation.
-
Configuration design
Constraint-driven enumeration of jumper geometries, batch prediction in the cloud.
-
BiLSTM surrogate
Per-element unity from the 32 design inputs; first web app.
-
Finite-element campaign begins
Automated Abaqus sampling, build and post-processing of M-shape jumpers.
Working together
The tools are free to use for strength prediction. If you are an operator, EPC or consultancy with jumper, spool or riser problems, the same approach can be trained on your own design basis. Research collaborations on surrogate modelling and guided-wave inspection are also welcome.
Get in touch →Using the results
DeepUnity is research software. Predictions come from models trained on finite-element data within the ranges of the training campaigns; they are not a substitute for a code-compliant verification by a qualified engineer. If the tools help your work, please cite the corresponding paper listed on the research page.