Title: Upward lightning at tall structures: Atmospheric drivers for trigger mechanisms and flash type

Abstract: Upward lightning is much rarer than downward lightning and requires tall ($>100\,m$) structures to initiate it. While conventional lightning locations systems (LLS) reliably detect downward lightning they miss the substantial fraction of upward lightning flashes that consist only of a continuous current. Globally, only few specially equipped towers can detect them. The proliferation of wind turbines in combination with large damage from upward lightning necessitates a reliable estimate of upward lightning frequency and conditions under which they occur. These estimates are computed by combining direct measurements at the specially-equipped tower at Gaisberg mountain in Austria as target variable with covariates from LLS measurements and atmospheric reanalysis data (ERA5) in a conditional inference random forest machine learning model. The most important factor determining whether upward lightning will not be detectable by LLS is the absence of nearby (within 4 km) lightning activity. All atmospheric covariates combined are ten times less important. Atmospheric variables, on the other hand, reliably explain whether upward lightning is self-triggered by the tower or other-triggered by nearby lightning discharges. The most important factor is height of the $-10~^\circ$C isotherm above the tall structure: the closer it is the higher is the probability of self-triggered flashes. Two-meter temperature and the amount of CAPE are also important. The results are an important step towards a comprehensive risk assessment of lightning damage to wind turbines and other tall structures.