Resumo:
Surface roughness Ra is a critical quality indicator in dry turning of Ti-6Al-4V titanium
alloy, a material widely used in aerospace and biomedical applications. Empirical powerlaw
models remain the industrial standard but cannot represent interaction effects or
non-linear responses. Machine learning offers greater flexibility, yet its deployment is
constrained by small experimental datasets and the need for rigorous validation protocols.
Large language models represent an alternative predictive paradigm, potentially encoding
implicit knowledge of machining physics that may support quantitative prediction through
in-context learning alone. This work compares five predictive models, comprising two
conventional mathematical approaches (Power Law and RSM) and three machine learning
models (SVR, Gaussian Process Regression, and XGBoost), against three LLMs (Claude
Sonnet 4.5, GPT-4o-mini, and Gemini 3 Flash Preview) for Ra prediction in Ti-6Al-
4V turning. A Central Composite Design with k = 3 factors and two carbide inserts
of distinct nose radii yielded N = 38 observations. The conventional and ML models
were evaluated under nested LOOCV with 5-fold grid search where applicable. LLMs
were accessed via API at temperature zero under two prompting strategies, few-shot and
chain-of-thought, across three independent runs, with results reported as mean ± standard
deviation. Claude Sonnet 4.5 in few-shot mode achieved the best overall performance
(RMSE = 0.288 ± 0.016 μm, R2 = 0.985), outperforming SVR, the best trained model
(RMSE = 0.341 μm), by 16% without task-specific training or fine-tuning. The effect of
prompting strategy was model dependent: few-shot prompting outperformed chain-ofthought
for Claude and GPT, whereas Gemini CoT achieved a substantially lower mean
RMSE than Gemini FewShot but exhibited markedly greater run-to-run variability. This
variability was especially pronounced for Gemini CoT, whose RMSE ranged from 0.31
to 0.99 μm across identical executions. The results show that a capable LLM supplied
with well-structured experimental demonstrations can surpass optimised trained models
for Ra prediction without task-specific training, while also demonstrating that predictive
performance and stability depend strongly on both model architecture and prompting
strategy.
