PSI - Issue 73

Lenganji Simwanda et al. / Procedia Structural Integrity 73 (2025) 138–145 Author name / Structural Integrity Procedia 00 (2025) 000–000

144

7

(b)

(a)

(d)

(c)

(e)

(f)

Fig. 3 SHAP dependence plots illustrating nonlinear effects and interactions between meteorological variables: (a) 2 m air temperature against ground snow load, (b) 10 m wind speed against ground snow load, (c) surface pressure against ground snow load, (d) surface downward thermal radiation against snow depth, (e) 2 m temperature against 10 m wind speed , and (f) 10 m wind speed against 2 m temperature. 4.3. Limitations It is important to note that the models reflect the assumptions of the simulated dataset (SIMELT). Any limitations in the physical model—such as fixed thermal properties or omission of wind scouring—may be inherited by the ML models. For real-world deployment, retraining on observed roof snow load data is recommended to capture all relevant effects. This study focused solely on U = 1.0 W/m²K and no sliding; varying U-values would shift the balance between accumulation and melt. Lower U-values (better insulation) retain more snow, potentially increasing the influence of solar radiation. Preliminary tests with U = 1.5 and U = 2.0 (higher heat loss) showed similarly high performance, with lower predicted loads due to rapid melt. These scenarios suggest a shift in dominant features and merit further investigation, which is reserved for future work.