Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DSC curve fingerprints directly encode mechanical properties of aluminum alloys

Lukas Pichlmann, Samuel Studer, Aurel R. Arnoldt, Paul Oberhauser, Johannes A. Österreicher

Abstract

Differential scanning calorimetry (DSC) is a standard tool for studying precipitation and phase transformations in aluminum alloys, yet its relation to mechanical performance has so far remained mostly indirect. Here, we demonstrate that DSC curves themselves act as fingerprints that directly encode mechanical properties. Four representative 6xxx series alloys (Al-Mg-Si) were subjected to different natural and artificial aging regimens, followed by DSC heat-flow measurements and tensile testing. Machine learning models trained on the thermograms predicted yield strength, ultimate tensile strength, and uniform elongation in five-fold grouped cross-validation, with the best model (Lasso) achieving R^2 values of 0.93, 0.86, and 0.87 and mean absolute errors of 14.3 MPa, 11.1 MPa, and 1.5 percent, respectively. Leave-one-alloy-out evaluation with sparse calibration using anchor samples further demonstrated generalization across alloy chemistries. While direct prediction on unseen alloy data degraded performance substantially, inclusion of as few as one to two anchor conditions from the target alloy recovered predictive accuracy, approaching that of the standard cross-validation. Feature importance analysis revealed that the 230 to 270 C region, associated with precipitation of the primary hardening phase beta'', contributed most strongly to predictive accuracy, providing direct mechanistic validation of the model. These findings establish DSC as a diagnostic tool that can serve as a rapid proxy for mechanical property evaluation, enabling accelerated alloy screening, process optimization, and integration of thermal analysis into data-driven manufacturing.

DSC curve fingerprints directly encode mechanical properties of aluminum alloys

Abstract

Differential scanning calorimetry (DSC) is a standard tool for studying precipitation and phase transformations in aluminum alloys, yet its relation to mechanical performance has so far remained mostly indirect. Here, we demonstrate that DSC curves themselves act as fingerprints that directly encode mechanical properties. Four representative 6xxx series alloys (Al-Mg-Si) were subjected to different natural and artificial aging regimens, followed by DSC heat-flow measurements and tensile testing. Machine learning models trained on the thermograms predicted yield strength, ultimate tensile strength, and uniform elongation in five-fold grouped cross-validation, with the best model (Lasso) achieving R^2 values of 0.93, 0.86, and 0.87 and mean absolute errors of 14.3 MPa, 11.1 MPa, and 1.5 percent, respectively. Leave-one-alloy-out evaluation with sparse calibration using anchor samples further demonstrated generalization across alloy chemistries. While direct prediction on unseen alloy data degraded performance substantially, inclusion of as few as one to two anchor conditions from the target alloy recovered predictive accuracy, approaching that of the standard cross-validation. Feature importance analysis revealed that the 230 to 270 C region, associated with precipitation of the primary hardening phase beta'', contributed most strongly to predictive accuracy, providing direct mechanistic validation of the model. These findings establish DSC as a diagnostic tool that can serve as a rapid proxy for mechanical property evaluation, enabling accelerated alloy screening, process optimization, and integration of thermal analysis into data-driven manufacturing.
Paper Structure (24 sections, 2 equations, 8 figures, 4 tables)

This paper contains 24 sections, 2 equations, 8 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Materials and methods
  3. Alloy compositions and heat treatments
  4. DSC measurements
  5. Tensile testing
  6. Data preparation
  7. Preprocessing
  8. Feature engineering
  9. Dimensionality reduction of raw thermograms
  10. Model fitting and evaluation
  11. Hyperparameter selection
  12. Cross-validation protocol
  13. Generalization across different 6xxx alloys
  14. Feature importance analysis
  15. Results and discussion
  16. ...and 9 more sections

Figures (8)

  • Figure 1: Baseline-corrected DSC thermograms for the four aluminum alloys in various tempers. Each narrow line represents an individual measurement, while bold lines show the mean curve of each alloy. For better readability, the magnified view shows the 200–350°C region for the individual measurements of each alloy.
  • Figure 2: Cross-validated prediction performance (on yield strength measurements) as a function of the number of latent components for PLS regression and PCA. Shown are the average $R^2$ (right axis) and mean absolute error (MAE) (left axis) scores across five folds.
  • Figure 3: MAE-distribution for yield strength and uniform elongation as a function of the number of anchor conditions ($n_\mathrm{anchor}$) in leave-one-alloy-out cross validation. Scatter points show all possible combinations of anchor samples from the validation set.
  • Figure 4: Variable Importance in Projection (VIP) scores as a function of temperature depicted as shaded background. The background intensity highlights the most predictive region. In general, VIP scores exceeding a threshold of 1.0 are deemed significant. The most impactful region is shown to occur between 230--270°C.
  • Figure 5: Lasso regression coefficients for the top 5 PLS components across the three mechanical properties. UE coefficient values are multiplied by 10 for better readability.
  • ...and 3 more figures