PINN TEAM LEAD
I lead the Physics-Informed Neural Networks team at ZLab. Our flagship project: 3-D heat-transfer simulation of injection-mold cooling, a geometry-aware network trained on a hybrid of simulation data and physics loss, and an optimization built on it that shortens cooling time and makes every region of the part cool together.
Lead the team, write the code, own the physics
ZLab is FANAP's Advanced Technologies Lab. I lead its PINN team, combining technical leadership with hands-on development and scientific research.
The mandate is to make engineering simulation faster and more useful with neural networks, without losing physical correctness. That is three jobs at once: setting the team's technical direction and research agenda, building a real share of the models and software myself, and making sure every result is checked against the physics before anyone relies on it.
- PINNsDesign and develop physics-informed neural networks for complex engineering and physics problems.
- Faster simulationResearch and implement network architectures that accelerate numerical computation and simulation.
- Geometry optimizationDesign and develop optimization algorithms for engineering geometry.
- Web platformsBuild web platforms that bring these AI solutions into engineering workflows.
- Applied researchApply current AI research to industrial problems for faster, more efficient simulation.
- Lead and buildLead technical development and research while contributing directly to the software.
READ MORE · HOW THE TEAM WORKS
Neural surrogates fail quietly: a training loss can look perfect while the model has learned something physically wrong. The team's process is built around catching that before it reaches a design decision.
Physics first
The formulation is settled before a network is touched: what is conserved, what is prescribed at the boundaries, and how the variables are non-dimensionalised.
Validated, not trusted
A prediction counts only after comparison with the reference solver on geometries the model never saw, not on the training loss alone.
Optima are re-checked
Any design the network proposes as optimal is confirmed by simulation before it is reported. The surrogate searches; the physics decides.
The part cannot leave until it is cold
In injection molding, molten plastic fills a steel cavity and must cool until it is rigid enough to eject. Cooling is usually the longest phase of every cycle.
Two things are decided by how the heat leaves the part. Cooling time: the part can only be ejected when its slowest region is cold, so one thick feature far from the coolant holds up the entire cycle. And cooling quality: regions that cool at different rates shrink at different rates, and that mismatch is locked into the part as residual stress and warpage. Straight drilled channels, the conventional layout, are placed where a drill can reach, not where the part needs them.
Fill, pack, cool, eject. The cooling block is drawn long because it usually is, and it is the phase a better cooling layout shortens.
READ MORE · THE PHYSICS
Cooling is transient heat conduction through two very different materials, the polymer and the mold steel, with heat finally removed by the coolant in the channels.
Initial state
The cavity is filled with melt at processing temperature; the steel starts near its operating temperature from the previous cycle.
Coolant
Channel walls exchange heat with the flowing coolant by convection, which sets how much heat each channel can pull out of the steel around it.
Ejection
The part may be ejected once every point is below the ejection temperature, so the cycle is set by the hottest point, not the average.
Why geometry dominates
The polymer conducts heat far worse than the steel, so the time a region needs grows quickly with its thickness and with its distance from the nearest channel. Thick ribs, bosses and inner corners are the usual hot spots. A layout that ignores them leaves the thin walls cold early while the hot spots set the cycle time.
Where residual stress comes from
Each region freezes and shrinks as it cools. When neighbouring regions cool at different times, the region that freezes last is constrained by the one already rigid, and that mismatch remains in the part as residual stress after ejection, showing up as warpage or reduced strength. Uniform cooling removes the mismatch at its source.
A network that reads geometry
We ran 3-D heat-transfer simulations of mold cooling, then trained a neural network on them that takes the geometry itself as its input.
A full 3-D simulation is accurate but too slow to call thousands of times inside a design search. So the network was trained with a hybrid loss: a data term that matches simulation results, and a physics term that penalises any violation of the heat equation. As the models improved, we shifted the weight increasingly onto the physics term, which let us train with a very small dataset of simulations. The trained network evaluates a new design almost instantly, which makes it the engine of the optimization: the search proposes cooling layouts, the network predicts how each one cools, and the most promising designs are confirmed with the full simulation.
Inputs cycle through three representative parts. Outputs: the cooling field, the residual-stress picture, and the cooling time. The bar shows the training loss shifting from simulation data toward the physics residual. Schematic of the approach, not the trained model.
READ MORE · METHOD
The pipeline separates the expensive, trusted physics from the fast, learned model, and lets each do what it is good at.
Why the physics term matters: a small dataset
Every 3-D simulation is expensive, so a model that needs thousands of them is not practical. We trained with a combination of both sources of information. The data term anchors the network to simulation results; the physics term checks the prediction against the heat equation itself, everywhere in the domain, with no simulation needed at those points. Early in training the data term does most of the work. As the models improved, we moved progressively more of the weight onto the physics term, and that is what made it possible to reach a reliable model with a very small simulation dataset.
Input · geometry
The network is given the geometry directly, so one trained model covers different parts and different cooling layouts instead of one fixed design.
Training · data + physics
Simulation data anchors the model; the heat-equation residual supervises it everywhere else. The balance shifts toward physics as the model improves.
Output · thermal response
How the part cools: the temperature field over time, the time until ejection, and how evenly the regions cool, which is what drives residual stress.
From model to optimum
Because evaluation is fast, the optimizer can explore far more cooling layouts than simulation alone would allow. Its best candidates are then handed back to the full 3-D solver, and only a confirmed design is accepted.
Scope of this page: the model, dataset, and industrial details of the project are not published here. The visuals are schematic and explain the approach and the kind of result it produced.
Shorter cooling, and the whole part cools together
After optimization the cooling time dropped, and the quality of cooling rose: every region of the part reaches ejection temperature at nearly the same moment, so no concentrated residual stress is locked into it.
Drag across the model to compare. On the left, the conventional mold: straight channels drilled through the plates. On the right, the optimized mold: a layout shaped by the part's geometry, running close to the floor, around the walls, and around the thick features. Switch the view between the cooling channels, the cooling profile at the moment the optimized part is ready to eject, and the residual stress that each layout leaves behind. Three representative parts are included.
READ MORE · WHAT CHANGED
The optimized layout does not simply add channels. It puts cooling where the geometry needs it, so the regions that used to lag no longer set the pace of the cycle.
| ASPECT | CONVENTIONAL | OPTIMIZED |
|---|---|---|
| Channel layout | Straight lines drilled through the plates, independent of the part's shape | Shaped by the geometry: close to the floor, around the walls, around thick features |
| Cooling time | Set by the hot spots far from any channel | Shorter: the slowest region is no longer far behind the rest |
| Cooling profile | Thin regions finish early, thick features and inner corners lag | All regions reach ejection temperature at nearly the same time |
| Residual stress | Concentrated where cooling was uneven: bosses, ribs, corners | Low and evenly spread, with no concentrated stress left in the part |
| Consequence | Longer cycle, higher risk of warpage | Shorter cycle, more stable part dimensions |
Why uniform cooling matters as much as fast cooling
A faster cycle that leaves the part distorted is not an improvement. The objective therefore weighed both: the time until the whole part can be ejected, and how evenly it gets there. Because residual stress comes from neighbouring regions freezing at different times, a profile where every region cools together removes the cause rather than treating the symptom.
About these visuals: the parts, layouts and fields shown are schematic representations prepared for this page, so the comparison can be explored without publishing the project's models or data.
Pay for the physics once
A classical solver pays the full cost of the physics every time it is asked a question. In design work it gets asked thousands of times.
That is the trade the mold-cooling network changes, and it is the same trade I made before in my airfoil research. The expensive physics is paid for once, in simulation and training, and after that the network answers in a fraction of the time. Physics-informed training keeps those fast answers tied to the governing equations instead of to whatever the training data happened to cover. The airfoil model runs live on this site: a geometry-aware PINN trained on 52,480 NASA CFD cases that predicts lift, drag and moment in about a millisecond.
The same idea, running live: the airfoil PINN on this site. R² ≈ 0.999 on airfoils it never saw in training, about 1 ms per prediction.
RUN THE AIRFOIL PINN →No solution data. Only the equation.
The physics term is what let the mold-cooling model work with so little simulation data: a physics-informed network is penalised whenever it breaks the governing equation itself.
This panel shows that term on its own, taken to the limit. It trains live on the same physics as mold cooling, heat conduction, in one dimension: ut = α uxx. The only data it sees is the initial temperature profile. Everywhere else the loss is the PDE residual at scattered collocation points. Watch the residual fall and the error against the exact solution disappear. Change the diffusivity and it re-learns from scratch.
To stay instant in a browser the model is a small spectral expansion trained with Adam, not a deep network, but the loss is exactly the physics-informed one.
A model nobody can use is not finished
A trained network in a notebook helps one person. The value arrives when an engineer can use it inside the work they already do.
So I also develop the web-based platforms that carry these AI solutions into engineering workflows: the model is wrapped as a service, given an interface an engineer can drive without knowing how it was trained, and connected to the design steps around it. The expensive physics is paid for once, in simulation and training; after that every query is fast.
- Simulate3-D physics-based runs across geometries and layouts.
- LearnTrain the geometry-aware network on a hybrid of simulation data and physics loss.
- VerifyCheck predictions against the solver on unseen designs.
- OptimizeUse the network as the fast engine of the design search.
- DeliverShip it as a web platform inside the engineers' workflow.
Deep technical range, and a team to lead
Leading a PINN team means being able to go to the bottom of the maths with anyone on it, and then keeping the whole team pointed at the same result.
// HARD SKILLS
- Physics-informed neural networksResidual losses, boundary and initial conditions, hybrid data + physics training.
- Heat transfer & numerical methods3-D transient conduction, finite-volume and finite-element thinking, non-dimensional scaling.
- Deep learning · PyTorchGeometry-aware architectures, training schedules, loss balancing.
- Surrogate modelling & optimizationUsing a learned model as the engine of a design search, then verifying the optimum.
- ValidationTesting against the solver on unseen geometries, not only the training loss.
- Python & web platformsPackaging models as services with interfaces engineers can use.
// SOFT SKILLS
- Team leadershipSetting direction, breaking problems into work people can own, reviewing results.
- Technical mentoringHelping teammates through the formulation, not only the code.
- Research planningChoosing which ideas to test, in what order, and when to stop.
- Communicating with industryExplaining what a model can and cannot promise to people who will rely on it.
- Judgement under uncertaintyDeciding when a result is good enough to act on.
The physics decides, the network makes it fast
The airfoil model on this site and the mold-cooling work at ZLab follow one approach: let simulation and the governing equations define what a good answer is, then use a geometry-aware neural network to deliver that answer fast enough to design with.