Physics-Aware Q-Learning for Adaptive Control of a Lubricated Sliding Contact under Variable Duty Cycles
International Conference on Tribology and Surface Science (TRiboSurf’26), Trabzon, Türkiye, 4 - 07 Kasım 2026, (Yayınlanmadı)
- Yayın Türü: Bildiri / Yayınlanmadı
- Basıldığı Şehir: Trabzon
- Basıldığı Ülke: Türkiye
- Karadeniz Teknik Üniversitesi Adresli: Evet
Özet
Tribological contacts operate under duty cycles in which load, sliding speed, temperature, and surface roughness evolve over time, limiting the effectiveness of fixed operating strategies. Because these variables are coupled through lubrication, friction, thermal, and wear dynamics, effective control requires sequential decisions that adapt to the evolving contact state rather than relying on predetermined settings. This study numerically investigates a physics-aware reinforcement learning framework for the adaptive control of a lubricated steel–steel sliding contact. A physics-based surrogate couples an elastohydrodynamic film-thickness relation, a load-sharing friction model, a lumped thermal balance, and an Archard wear model with wear-induced roughness evolution. A tabular Q-learning agent adjusts the normalized lubricant supply and speed trim at two-second intervals. Its reward balances dissipated power, incremental wear, throughput loss, and excessive setpoint variation. A one-step action-filtering layer screens candidate actions using the predicted film parameter and contact temperature, thereby restricting the agent to actions that satisfy the nominal operating constraints whenever feasible. The agent was trained on randomized multistage duty cycles and evaluated on 40 held-out cycles against fixed-flow strategies, a rule-based controller, and a model-based one-step optimizer. Relative to rule-based control, the constrained policy reduced dissipated energy by 16.4% and decreased constraint-violation time from 3.33% to 1.83%, while the mean coefficient of friction and wear volume increased by 13.8% and 16.9%, respectively. The model-based one-step optimizer achieved a higher mean reward, whereas the learned policy produced 4.5% less wear, consistent with the intended influence of delayed wear–roughness feedback on sequential decisions. Compared with unconstrained Q-learning, action filtering reduced the violation rate from 2.34% to 1.83% with essentially unchanged mean reward. These results demonstrate the simulation-based potential of physics-aware reinforcement learning for managing energy–wear–throughput trade-offs under variable operating conditions. The conclusions remain limited to the nominal surrogate model and require verification under model uncertainty and experimental conditions.