Physics-based models
are well understood.
Can new architectures
push them further?
I combine first-principles vehicle dynamics with machine learning — because a model should not just be accurate, it should be understandable. Physics provides structure. Data corrects what the equations miss.
from Autumn 2026
Predictive
State
Estimation
How do physics-based, data-driven, and hybrid vehicle models compare for real-time-capable predictive state estimation? Which approach offers the best trade-off between prediction accuracy, computational cost, and interpretability?
Trained on rig measurement data. Evaluated autoregressively. Metric: prediction accuracy vs. real-time compute budget.
Physics vs. Data
vs. Hybrid
Three model classes on real rig data — which offers the best accuracy-to-compute trade-off?
Models break
at the limit
STM/DTM lose accuracy near the friction limit due to linearisation. Data-driven models generalise poorly outside training data. Real-time MPC needs accuracy, speed, and explainability simultaneously.
Three-way
comparison
Physics-based (STM/DTM + Pacejka MF), data-driven (neural net / GP), and hybrid (physics + residual ML) — all trained on rig measurement data, evaluated autoregressively over a prediction horizon.
Hybrid wins
accuracy + cost
Prediction
Accuracy
vs. Cost
Three model classes evaluated autoregressively on rig measurement data across a defined prediction horizon. States: position, velocity, lateral acceleration.
| Model | RMSE ↓ | R² ↑ | MAE ↓ | Compute ↓ |
|---|---|---|---|---|
| STM Baseline | — | — | — | Ref. |
| DTM Baseline | — | — | — | — |
| Data-Driven | — | — | — | — |
| Hybrid | — | — | — | TBD |
UVA
Link Lab
STM vs. DTM
at the limit
Dual-Track Model with Pacejka MF validated on IAC telemetry — why kinematic models fail at 1.8 g.
STM fails
at IAC speeds
The Dallara AV-21 operates at 270+ km/h and 1.8 g lateral. STM linearises tyre forces — accuracy degrades exactly where it matters most, near the friction limit in high-speed cornering.
DTM with
Pacejka MF
Derive a full Dual-Track Model with load transfer and Pacejka Magic Formula 2002 tyre model in MATLAB. Validate against Cavalier Autonomous Racing IAC telemetry at UVA Link Lab.
DTM selected
for thesis & IAC
Trajectory
Comparison
Comparison of Single-Track and Dual-Track model prediction accuracy for the Dallara AV-21. Real IAC telemetry is proprietary — conceptual illustration only. DTM selected as basis for master thesis work.
Conceptual illustration. Real IAC measurement data is not publicly shareable.
Linear. Good at low-lateral. Loses accuracy at limit due to linearisation.
Captures load transfer & tyre nonlinearity. Better at lateral limits.
| Model | RMSE aᵧ ↓ | Max err ↓ | Δ vs. STM |
|---|---|---|---|
| STM Baseline | — | — | Ref. |
| DTM | — | — | −XX% ↓ |
Vehicle
Dynamics
Tool
Modular simulation framework — built out of intrinsic motivation. Fed by every lecture, course, and day trackside. No institutional context.
Growing modules: STM → DTM, Pacejka MF, brush model, load transfer, QSS lap simulation, GGV envelope.
Modular VDT
from scratch
No institutional backing — built from every lecture, course, and day trackside.
No accessible
modular VDT
Professional tools (CarSim, ADAMS) are closed and monolithic. No lightweight framework for rapid prototyping of new model formulations — the gap between lecture theory and real trackside dynamics is wide.
Modular MATLAB
from scratch
Self-directed framework — each physics block independent and swappable. Built incrementally from every lecture, course, and day trackside. No institutional backing.
Working sim
+ browser GUI
Pacejka MF · QSS
VDT Browser GUI
GUI only — no physics backend yet. Parameter sweeps and real-time plot visualisation.
F1TENTH
Sim Race
Full autonomous stack for the F1tenth 1/10-scale racing platform. Builds on IAC experience. First target: ICRA Vienna 2026 simulation race.
From IAC
to F1TENTH
Porting thesis vehicle models into a ROS2 autonomous stack for 1/10-scale racing.
Scale gap:
AV-21 to 1/10
IAC experience from the 290+ km/h Dallara AV-21 doesn't transfer directly to the 1/10-scale F1TENTH. Standard stacks use kinematic models — insufficient for performance racing near the friction limit.
full scale
sim race
Thesis models
into ROS2 stack
Port STM/DTM from the master thesis into the F1TENTH ROS2 stack. Full pipeline: LiDAR → state estimation → MPC with physics plant model. Simulation-first, hardware later.
Post-thesis
in development
Race
Performance
DTM-MPC vs. kinematic baseline on the ICRA qualification map. Physics-informed plant model shows faster lap time convergence and higher top speed.
Design
Projects
CAD and surface design work running parallel to simulation. F1 sidepod study in CATIA V6 as part of MEA CAD CAP. PoliMi Roboracer surface model. FSAE cockpit in Creo 5 — −70% weight via generative design.
CATIA V6
MEA CAD CAP
External aero surface study — full sidepod geometry including inlet, undercut, and fin. Final project of the MEA CAD CAP course. Advanced Class-A surfacing workflow in CATIA V6.
Creo 5
−70% Weight
FSAE cockpit redesign at GETracing Dortmund. Full CAD redesign of dashboard, steering wheel, and shift paddles. −70% weight reduction vs. original via topology optimisation and generative design workflow in Creo 5. Validated through real-driver fit testing.
PoliMi
Surface Model
Surface model and render produced at Politecnico di Milano as part of the Surface Design for Engineering Applications course. Advanced Class-A surface modelling for the PoliMi Roboracer autonomous platform.
Universities
↑ here
Present
Sep 2022
Practical
Experience
Apr 2024
May 2022
Sep 2019
Jul 2023