Physics-informed interaction modeling for warning targeting in multi-risk driving scenarios

In multi-risk traffic, a warning is valuable only if it directs the driver's limited attention to the right interaction counterpart early enough to preserve feasible evasive actions. Existing proximity- or surrogate-metric-based warnings capture direct geometric conflicts but often miss behavior-mediated hazards that become critical through the driver's own evasive maneuvers. We propose a physics-informed framework that makes ego-agent interaction explicit and quantifiable. A neural network decodes time-varying inter-agent coupling and control sensitivity from scene history and map context, yielding an interpretable interaction-strength signal for each surrounding agent. A warning is triggered when a two-sample Kolmogorov-Smirnov test detects an abrupt distributional change in interaction strength, and is anchored to the agent with the strongest intensification. Critical-agent identification reaches 83.07% agreement with a physically interpretable baseline on 3131 naturalistic scenarios from the Waymo Open Motion Dataset. In a driver-in-the-loop virtual reality study (n=31) within a multi-risk scenario, the proposed interaction-aware human-machine interface (HMI) elicits faster attention shifts than an anticipated post-encroachment time (APET)-based HMI (head-turn responses within 0.2s versus a 0.8s lag) and reduces the observed collision events from 12 (no-warning) and 7 (APET-based HMI) to 1. These results indicate that quantifying interaction strength to identify critical counterparts offers a practical route to warning targeting in complex multi-risk driving.