Nicolas Smits

SiNDAE

CAPD, Carnegie Mellonwith Laurens Lueg and Victor AlvesJun – Aug 2026

h₁, tank 1 levelh₂, tank 2 level
A₁ḣ₁ = Fᵢₙ − q₁   A₂ḣ₂ = q₁ − c₂√h₂   q₁ = NN(h₁, h₂)   The network learns the discharge q₁; the valve law of tank 2 stays a hard constraint. Peach: the three training trajectories from the paper's example; coral: the steady state.

Motivation

Mechanistic models of process systems are often incomplete: a reaction rate, a growth term or a transfer coefficient depends on the states in a way nobody has written down. A hybrid model keeps the known balances and learns the missing term from data, so it stays physically consistent where the data run out. Open-source tools for this mostly target ODEs and train by integrating and backpropagating; few handle differential-algebraic systems, where algebraic constraints must hold exactly alongside the dynamics.

Method

The network enters the model as one more algebraic equation, and the whole system is discretised by Lagrange–Radau collocation:

𝐱˙=𝐟(𝐱,𝐲,𝐳,𝐩),0=𝐡(𝐱,𝐲,𝐳,𝐩),0=𝐳−fNN(𝐯,𝜽)

The weights and the trajectories are then solved for together, so the constraints hold at every collocation point by construction, index-2 systems included. The decomposition backend instead updates the weights in an outer Adam loop around per-trajectory solves, with gradients from the KKT system, and spreads across trajectories with MPI.

Two-tank example

Two tanks in series under a constant feed, the system in the figure at the top of the page. The outlet valve law of the second tank is known; the discharge between the tanks is not, and is learned from noisy level measurements alone. The network is given no structure, yet its response falls onto the true law, and predictions from unseen initial levels track the reference.

Two tanks in series: feed into tank 1, discharge q1 through a valve into tank 2, outflow q2
The two-tank cascade. The discharge q₁ between the tanks is the unknown term.Figure from the JOSS paper.
Learned discharge against tank 1 level, lying on the true square-root law
The learned network against the true law c₁√h₁, from noisy level data alone.Figure from the JOSS paper.
Tank levels and discharge over time for three trajectories: ground truth, a smoother and the hybrid fit
The hybrid fit against ground truth and a smoother, on three trajectories.Figure from the JOSS paper.

Package

A scikit-learn-style HybridDAE runs smoothing, pre-training, training and prediction behind fit and predict. The default install is pure pip wheels, with the POUNCE and FERAL solvers replacing licensed HSL binaries, and trained networks export to ONNX, Equinox, OMLT and JSON. A sindae-skills bundle lets Claude Code draft and check a hybrid model with the user before any code runs.

Further reading