CALIBRATED DIGITAL TWINS AND GEOMETRY-BASED PERCEPTIONFOR AUTOMATED DRIVING IN HEAVY RAIN AND SNOWFALL:EVIDENCE FROM RESEARCH TESTING IN CHINA AND IMPLICATIONSFOR KAZAKHSTAN

Аңдатпа

Automated driving systems that perform reliably under clear conditions exhibit substantial

degradation in heavy rain and snowfall. Precipitation attenuates and scatters LiDAR returns

and generates spurious near-range points, water and spray impair camera imagery, whereas

4D radar has proved more robust than LiDAR in fog, rain and snow. Validation through

public-road driving alone is impractical for these conditions, because severe weather events

are infrequent, seasonal, hazardous for test personnel and not reproducible. This study

examines whether a calibrated digital twin, understood as a virtual representation of the

vehicle, its sensor suite and the meteorological environment that is iteratively corrected

against physical measurements, constitutes a necessary component of severe-weather

validation. The analysis is based on a structured narrative review of experimental testing

conducted in China between 2009 and 2026, including the Intelligent Vehicle Future

Challenge, virtual–real parallel testing, closed test fields that simulate rain, fog and other

weather in Changsha, Wuhan and Chongqing, the cold-region proving grounds of Heihe, and

recent studies that calibrate simulated rainfall and snowfall noise against field measurements.

The results indicate that each testing modality addresses only part of the validation problem.

Closed-field rainfall systems are repeatable but spatially non-uniform, so that nominal rainfall

intensity overstates test realism; cold-region proving grounds provide natural snow and ice

but are available only from November to March; and simulated rainfall is scalable and fully

controllable but becomes credible only when parameterised with measured drop-size

distributions and sensor responses. A digital twin that integrates these modalities converts

scarce physical tests into calibration points and enables year-round exploration of the scenario

space. The study further specifies a geometry-first machine-learning layer within the twin,

supported by camera and LiDAR models adapted to heavy rain and snow, that infers the

drivable corridor from road-surface elevation, road edges, snow banks, the tracks of preceding

vehicles and three-dimensional occupancy rather than from painted markings or road signs,

and that triggers a minimal-risk manoeuvre when geometric confidence is insufficient or a

road is officially closed. The framework is applied to Kazakhstan using data from RSE

Kazhydromet, which recorded 190 hazardous meteorological events in 2024, including 28

severe blizzards with visibility as low as 50 m, and which reports road closures among the

consequences of strong winds. The study proposes a weather transfer gap, G = |Preal − Ptwin| /Preal, as an acceptance criterion for substituting virtual for physical evidence, and specifies the

data that test facilities and regulators would need to report for its application.

Keywords: automated driving; adverse weather; heavy rainfall; snowfall; digital twin;

parallel testing; closed-field testing; LiDAR; sensor simulation; scenario-based validation; 3D

occupancy; road closures; Kazakhstan.