Research

Offshore multiphase pipelines leak, and detecting a leak, sizing it and locating it are usually three separate problems, solved either by transient models too expensive to run in the loop or by data-driven detectors with no physical consistency. PCE-PINN does all three at once, from inlet and outlet pressure and flow rate alone. A physics sub-network penalises violations of nondimensional mass, momentum and mechanical-energy balance through adaptive curriculum training and adaptive collocation.

Finding. Detection is saturated at F1 = 1.00. The regression heads are what matter: location R2 improves 16.5 percent over the best of five classical baselines on identical features, so the gain comes from the physics-informed architecture rather than from the features. SHAP attribution lands on conservation-consistent variables.

Authors. Mridha, Kabir, Khandakar, Mia, Ahmed, Shahadad, Jim, Rahman, Ayari, Rahman and Barooah. Departments of Electrical and Electronic Engineering at the University of Dhaka and Qatar University, with Hamad Bin Khalifa University.

A single forward solve of a high-resolution two-phase CO2 reservoir simulation takes hours, and history matching needs thousands of them. Embed-to-Control surrogates compress the state into a latent vector and advance it under well controls, but they share one linear observation model across every well, which regresses water and gas rates with very different non-linearities into one mediocre fit. PINN-E2CO replaces that block with an independent MLP head per producer over the shared latent, warm-started from a physics-free baseline at a tenth of the head learning rate, and trains under a hybrid loss whose three physics residuals are discretised on the finite-volume transmissibility operator the simulator itself uses.

Finding. Per-well heads are the fix, and the physics residuals are what make them stick: saturation RMSE down 38 percent and pressure RMSE down 35 percent against the non-physics baseline, with filtered per-well R2 moving from 0.82 to 0.91 after 16 minutes of fine-tuning on a single T4. An ablation of 33 variants, including FNO decoders, neural-ODE dynamics and DeepONet heads, confirms the shared linear observe block was the bottleneck.

Authors. Khandakar (corresponding), Ruhan, Kabir, Jim, Mia, Ahmed, Rahaman, Rahman, Ayari and Rahman.

The AI4Pain 2026 challenge asks a model to label physiological segments from 65 subjects as No-Pain, Pain-Arm or Pain-Hand from BVP, EDA, respiration and SpO2. We present a two-stage hierarchical classifier: stage one separates pain from no-pain, stage two separates arm from hand. Both are L2-regularised logistic regressions on subject-normalised, effect-size-ranked features with isotonic calibration, decoded under the known balanced prior. Every design decision was made under 5x5 repeated subject-fold cross-validation with paired significance testing, not on the leaderboard.

Finding. Stage-one detection is strong. Stage-two arm-versus-hand localisation sits near chance and stays there through baseline-relative normalisation, respiratory dynamics, modality late-fusion, a sweep of non-linear models and an oracle variant fed ground-truth labels. At the segment level these peripheral signals carry pain intensity, not somatotopic location.

A CNN2D-TCN with late fusion of statistical EEG features classifying seven motor tasks across sixteen subjects under mixed leave-one-subject-out evaluation.

Finding. The release task (T6) is the stubborn one at 68 to 78 percent F1 in every configuration tried; everything else is well above 85.