US bridge inspectors now must track dozens of tiny damage points instead of one rating — forcing a complete rethink of maintenance schedules

US bridge standards changed in 2022 to require inspectors to measure the condition of individual bridge elements (like girders, joints, bearings) as probability distributions across multiple damage states, rather than a single overall rating. This makes inspection data vastly more detailed, but it broke the old computer models for deciding when to repair or replace bridges, since those models were built for simple categorical ratings and couldn't handle multi-dimensional probability arrays.

Bridge managers have been making maintenance decisions using decades-old rules of thumb designed for simple condition ratings. The new standard gives them much richer data about what's actually wrong with each bridge component, but now they have no proven way to turn that detailed information into repair decisions. This paper offers one solution: a machine learning system that learns optimal repair strategies directly from element-level condition data and outputs its decisions as simple decision trees that human engineers can read and audit. That matters because bridge agencies don't trust black-box AI — they need to understand and defend their spending choices. If this approach works at scale, it could squeeze more life out of aging bridge stock and make repairs more targeted and cost-effective. If it doesn't, bridge managers are stuck with granular data they can't actually use.