ERHAM — Aviation Application & Research Center · Erciyes University · Türkiye

Machine Learning–Assisted Digital Twin for Aircraft Conceptual Design MDO

A full optimization cycle in 5 minutes instead of 1.2 hours, with each prediction's uncertainty stated up front.
Elif Zeyneb Tekin · Melih Yıldız
16th EASN International Conference
Toulouse · 27–30 Oct 2026 · Digital Design session

The Framework: One Loop, Two Speeds

Parametric Aircraft Definition 6 design variables DOE Sampling LHS · 140 points Physics-Based Sizing Loop 431 ms / point Aero- dynamics Weights (MTOW loop) Hybrid Propulsion Mission Performance Pareto Front block energy per seat-km · take-off mass verified against the physics model training data (248 physics runs in total) ⟳ update Pareto-guided infill: new physics runs only where the front is uncertain candidate front The design iterates at surrogate speed; physics runs only where the twin is uncertain. PREDICTIVE DIGITAL TWIN Gaussian Process Surrogates prediction ŷ + stated uncertainty σ Matérn 5/2 · 2 objectives + 8 constraints 51 µs / point NSGA-II Optimizer ĝ + κσ ≤ 0 uncertainty-gated constraints query x ŷ ± σ

The Challenge

Conceptual design locks in most of an aircraft's life-cycle cost, yet it is the stage where physics-based analysis is used least, because each design point costs too much. Hybrid-electric and hydrogen configurations have no historical data to fall back on, so their design space has to be explored with physics.

Three Findings

  1. Judge a surrogate where the optimizer uses it. Active learning gives no global accuracy gain over an equal-budget static sample (0.50 % vs 0.48 %), but at the Pareto front its error is 29× lower (0.03 % vs 0.84 %).
  2. −10.5 % block energy, fuel and CO₂ relative to an ATR 72‑500 reference that the model reproduces to within 1 %. The gain comes from aspect ratio 15.9 and cruise L/D rising from 14.9 to 16.9.
  3. Battery specific energy is the gate. At 350 Wh/kg (pack level) the energy optimum is a conventional turboprop. Hybridisation starts to pay above 560 Wh/kg and saves 13.7 % at 1500 Wh/kg.

Measured Cost — 70-seat hybrid-electric regional aircraft

Physics modelThis framework
One design evaluation431 ms51 µs ×8 400
Full optimization cycle, all training charged1.16 h5.0 min ×14
Re-run after a changed requirement1.16 h44 s
Mean error, 120 independent designs—0.50 %
Error at the Pareto front0.84 % (static)0.03 % ×29

The physics layer is mid-fidelity, so these speed-up ratios apply to it only. With RANS/FEM the per-point ratio grows, while the number of samples one can afford shrinks. The framework itself is fidelity-agnostic.

We are looking for: an industrial use case to test the framework against a real early-stage trade study, and partners for Horizon Europe / Clean Aviation proposals.

Companion talk at this conference: Experimental Validation of an Aircraft Digital Twin: From Concept to Validation (A. N. Dişlitaş, M. Yıldız), which asks how good a test must be before it can validate a model like this one.