02TÜBİTAK 2209-A · Research

Reading a materialfrom its heat.

The same laser pulse leaves six different heat traces on six materials. A model that recognises them can support choosing the energy per material.

Programme: TÜBİTAK 2209-A · funded, completed
My role: Project lead · ML / CV
Model: DualViT-CNN + ArcFace
Publication: Gazi Univ. J. Sci. Part A · 2026

My contribution

I was the project lead; we were a team of three with one advisor. The machine learning and computer vision work was mine: data preparation and preprocessing, designing and training the DualViT-CNN + ArcFace model, and comparing and evaluating it against classical models. The ANSYS simulations were team work.

02Explanatory architectureTÜBİTAK 2209-A

The homepage drawing: the input is the report’s original thermal image; the architecture is redrawn for explanation.

01Problem

One pulse, six different traces.

When PCBs are processed with a femtosecond laser at one fixed energy density (fluence), one material receives more energy than it needs and another too little. Materials carry heat very differently: their thermal conductivities differ by about four orders of magnitude.

The study turns the thermal response of aluminium, copper, FR-4, graphene, PTFE and titanium under the laser into images by simulation, and builds a model that recognises the material from that image. Once the material is known, the energy can be chosen for it.

Two similar traces

FR-4 and PTFE conduct heat almost equally poorly (0.29 and 0.25 W/m·K); in both, the trace stays in a small spot. They are also, more often than not, the pair the models confuse most.

Aluminium: 2D thermal image from the simulation — The report’s original images (Figure 4) and the thermal conductivities used in the simulations (Table 2, log scale). With high conductivity the heat spreads; with low conductivity it stays in a point.

Copper: 2D thermal image from the simulation — The report’s original images (Figure 4) and the thermal conductivities used in the simulations (Table 2, log scale). With high conductivity the heat spreads; with low conductivity it stays in a point.

02Data

From simulation to dataset.

A thermal analysis in ANSYS for every material and board standard, then one uniform image the model can read.

A tetrahedral mesh built in ANSYS Workbench for different board standards. — The mesh for the board standards in ANSYS (report, Figure 2).

Total heat flux in a steady-state thermal analysis of a titanium VME 6U board, with a heated region in its centre. — The mesh for the board standards in ANSYS (report, Figure 2).

The steady-state analysis set a constant 100 °C surface temperature and h = 10 W/m²K convection on the edges; the transient analysis modelled the laser pulse as a Gaussian heat source and solved 0–5 ms. Corrupted, saturated or misplaced outputs were removed: 736 images.

I prepared the images for the model: fitting them to 585 × 395 pixels with the aspect ratio kept and zero padding, converting to grey so the model could not memorise the colour map, and normalising to [0, 1]. For the classical models I extracted a 256-bin intensity histogram from every image.

Augmentation was applied to the training set only: horizontal flips and ±15° rotations, taking the set from 736 to 2,208 samples.

The ANSYS simulations were team work; turning the images into a dataset and preprocessing them was mine.

Six panels: surface temperature distributions on aluminium VME 6U, copper CompactPCI 6U, FR-4 PMC, graphene CompactPCI 3U, PTFE VME 3U and titanium PC104 boards, each with its pixel size. — Surface temperature distributions across board standards. Their different sizes are why images are fitted to one size (report, Figure 5).

The original thermal image of a graphene VME 6U board. — The graphene VME 6U image and two rotation examples (the three panels of the report’s Figure 8 embedded in the report).

The same image rotated. — The graphene VME 6U image and two rotation examples (the three panels of the report’s Figure 8 embedded in the report).

03Model

Two views, one signature.

A CNN for local texture, a Vision Transformer for the whole image, and a cross-attention layer that joins them.

03DualViT-CNN + ArcFaceRedrawn

Input224 × 224 thermal image

CNN · ResNet-18Local texture → query (Q)

ViTGlobal context → key, value (K, V)

Cross-attentionQ meets K, V

Average poolingOne representation vector

ArcFaceAngular margin

Six materialsAl · Cu · FR-4 · Graphene · PTFE · Ti

An explanatory diagram redrawn from the report’s description of the architecture (Section 2.4, Table 9).

The CNN branch (ResNet-18, pretrained on ImageNet) captures the local temperature gradients and small textures the laser leaves. The ViT branch learns relations between distant parts of the image: how the heat spreads across the surface.

In the cross-attention fusion, CNN tokens act as queries and ViT tokens as keys and values, so a local hot spot is read together with the distribution over the whole surface. For classification, ArcFace adds an angular margin between classes, which helps separate close materials such as FR-4 and PTFE.

According to the report’s ablation summary, removing the cross-attention layer caused the largest drop, and using Softmax instead of ArcFace increased confusion between similar materials. The summary is not backed by repeated-run logs.

03Why ArcFace?Explanatory

Softmax

ArcFace

Explanatory drawing: ArcFace adds an angular margin to each class direction, pushing similar classes apart.

04Results

Where are the mistakes made?

Six models, two data set-ups. The hybrid model was the most accurate in both; in both, its mistakes fell on the same two pairs: PTFE and FR-4, aluminium and titanium.

04Model comparisonAccuracy · F1

Original data · 148 test images
Model	Accuracy	F1 (macro)
kNN	75.00%	74.84%
Decision tree	84.46%	84.65%
Bagging	89.19%	89.23%
Random forest	89.19%	89.26%
Gradient boosting	91.22%	91.12%
DualViT-CNN + ArcFace	95.95%	95.94%

Augmented data · 736 images
Model	Accuracy	F1 (macro)
kNN	94.70%	94.74%
Decision tree	97.15%	97.17%
Bagging	98.37%	98.36%
Random forest	98.64%	98.63%
Gradient boosting	97.83%	97.82%
DualViT-CNN + ArcFace	99.18%	99.17%

The report’s Tables 10 and 11. The classical models worked on 256-bin intensity histograms, the hybrid model on the images themselves. The hybrid’s rows match the published article’s abstract.

04Confusion matrixDualViT-CNN + ArcFace

Original data · 148 test images · Rows are the true class, columns the prediction
	Al	Cu	FR-4	Gr	PTFE	Ti
Al	20	0	0	0	0	2
Cu	0	25	0	0	0	0
FR-4	0	0	25	0	1	0
Gr	0	0	0	25	0	0
PTFE	0	0	3	0	22	0
Ti	0	0	0	0	0	25

Augmented data · 736 images · Rows are the true class, columns the prediction
	Al	Cu	FR-4	Gr	PTFE	Ti
Al	105	0	0	0	0	3
Cu	0	125	0	0	0	0
FR-4	0	0	126	0	0	0
Gr	0	0	0	125	0	0
PTFE	0	0	3	0	123	0
Ti	0	0	0	0	0	126

Transcribed from the report’s Figures 19 and 20. The diagonal holds correct predictions; coloured cells are mistakes.

With the original data, 6 of 148 images were wrong: 3 PTFE images taken for FR-4, 1 FR-4 image for PTFE, and 2 aluminium images for titanium.

With the augmented set-up, 6 of 736 images were wrong: 3 PTFE images taken for FR-4 and 3 aluminium images for titanium.

The two set-ups were evaluated on different sets (148 and 736 images). The augmented result should not be read as a direct improvement on the original one; images derived from the same simulation resemble each other.

The report’s original matrices

Every model’s confusion matrix in both data set-ups. Tap one to enlarge it.

kNN
Original Augmented
Decision tree
Original Augmented
Random forest
Original Augmented
Bagging
Original Augmented
Gradient boosting
Original Augmented
DualViT-CNN + ArcFace
Original Augmented

05Energy

The right material, the energy it needs.

The fixed approach processes every board at copper’s high threshold. Once the material is known, low-threshold boards need less.

The fixed reference fluence is 0.40 J/cm². With model-based selection, the average fluence set by the classified material is ≈0.209 J/cm² in the original set-up and ≈0.211 J/cm² in the augmented one: in theory, about 47% less energy.

The riskiest mistake is PTFE taken for FR-4, which can apply too little energy. The report proposes a simple feedback rule: if no effect is seen after an FR-4 prediction, one 0.30 J/cm² correction pulse is fired at the same spot. The estimated saving then falls from 47.2% to 46.9%.

05Fluence comparisonReport, Figure 21

Fixed reference

0.400

Model · original

0.209

Model · augmented

0.211

≈ 47% theoretical saving

The report’s values (Section 2.5, Figure 21), redrawn.

Energy saving per pass rises linearly with board area: from about 16 J for PC/104 to 70 J for VME/CompactPCI 6U. — The report’s original chart: the saving grows with board area, to about 70 J per pass on large boards such as VME 6U (theoretical).

The saving is a theoretical estimate based on simulation; there is no physical laser experiment or measured energy consumption. The article itself describes the approach as preliminary decision support.

06Publication

From a project to an article.

01
TÜBİTAK 2209-A
The project was funded and completed (8 April 2025 – 3 January 2026). I was its lead.
02
Publication
The work was published as an article in Gazi University Journal of Science Part A (2026).
03
Model
DualViT-CNN + ArcFace: 95.95% accuracy and 95.94% F1 on the original data.
04
Scope
Simulation-based classification and a theoretical energy estimate; physical validation is the next step.

Article

Özel Y., Vall M. M., Yıldırım M. A., Balcı H. Ş., Balcı F., Ilgın H. A. AI-Assisted Thermal Response Classification of PCB Materials for Energy-Efficient Femtosecond Laser Processing: A Simulation-Based Study. Gazi University Journal of Science Part A: Engineering and Innovation, 13, 1–19 (2026).

Team: three people and one advisor. The article has six authors; the author order follows the official record.

Academic record (AVESİS)Next: the search & rescue system Back to the research chapter

The report’s original matrices

TÜBİTAK 2209-A

Publication

Model

Scope

Let’s buildthe next system together.