DeepUnity builds fast, physics-faithful models for offshore pipe problems: neural surrogates that predict jumper strength and fatigue along every element, and a GPU-batched solver for guided-wave dispersion in pressurized pipes.
M-shape jumpers connect subsea wells, manifolds and pipelines, and every one is a bespoke structure. Surrogates trained on a large finite-element campaign return per-element strength and fatigue in seconds, so a design can be screened during the meeting that asked for it.
Ultrasonic guided waves are how pipelines are inspected, and internal pressure shifts their dispersion curves. GPU-SAFE is a semi-analytical finite element solver batched on the GPU, with exact gradients, for computing those curves across thousands of pipe geometries and pressures at once.
Jumpers are sampled across geometry, pipe section, pressure, temperature, water depth, hub tolerances and end expansions, then solved in Abaqus with nonlinear geometry.
The element chain is read like a sentence: type, position and the 32 global inputs at every element. A BiLSTM predicts unity directly; a transformer predicts axial force and bending moments.
Predicted forces become stresses by the same arithmetic the FEM post-processor uses: ASME B31.8 design factors for strength, DNV-RP-C203 S-N curves and Miner summation for fatigue.
At milliseconds per case the design tool enumerates every feasible configuration for a span and elevation, ranks by target unity, and reports where the governing stress occurs.
The models never see a stress they were not trained on. Everything downstream of the network is deterministic engineering arithmetic, so a prediction can always be traced back to a formula. The guided-wave project takes the other route: no surrogate at all, but a direct finite-element solver made fast enough on a GPU to sweep thousands of cases.
The BiLSTM surrogate behind the strength tool, and the constraint-filtered configuration search behind the design tool. Trained on about 150,000 finite-element models; validation RMSE about 0.07 unity.
A sectional-force transformer with DNV-RP-C203 post-processing, validated on a 1,000-run paired operating-state campaign.
A feature-importance and ablation study of the BiLSTM: end expansion and pipe section dominate, geometry follows, and position matters modestly.
Differentiable inverse design looks elegant; with manufacturing limits and hard constraints it becomes brittle. Why a forward-only, constraint-filtering workflow wins here.
DeepUnity is developed by Chi Yang, Ph.D., a versatile engineer working across offshore structures, machine learning and high-performance computing. It began with a simple question: if a FEED-level jumper design takes days of finite-element iteration, how much of that can a model trained on the same physics do in seconds?
The site is a research tool and a showcase. Strength prediction is free to use. If you are an operator, EPC or consultancy with jumper, spool or riser problems of your own, get in touch.
