### 2.442 Javan-Emrooz et al. (2018)

• Ground-motion model for horizontal PGA:
where PGAH is in cms2 and σ = 0.295 for training set and σ = 0.300 for testing set. Ground-motion model for vertical PGA:
where PGAV is in cms2 and σ = 0.295 for training set and σ = 0.315 for testing set.
• Use 2 site classes because of limited number of records:
1.
Rock/stiff soil, V s,30 375ms. S = 1.
2.
Soil, V s,30 < 375ms. S = 2.

Use method based on topographic slope to classify stations without measurements.

• Use 2 faulting mechanisms because only 9 records from normal-faulting events:
1.
Reverse, rake angles between 30 and 150. F = 1.
2.
Strike-slip, rake angles within 30 of horizontal. F = 0.25.
• Data from 1976 to 2016. Mainly from the Building and Housing Research Center (Iran) (419 records) and the Disaster and Emergency Management Presidency of Turkey (41 records) with 2 records from Armenia and 1 from Georgia. Records mainly from SSA-2 instruments (385 records) with some from SMA-1 (34), CMG-5TD (32), GSR-16 (6), SMACH (5) and SM-2 (1).
• Use an analysis of variance technique (Douglas2004b) to confirm that there is no significance difference in ground motions in Alborz-Azarbayejan and Kopeh Dagh regions. Find some significant differences between ground motions in Iran and Turkey/Armenia/Georgia so derive additional models using only Iranian data (not reported here due to lack of space).
• Only use data from Mw 4.5 to concrete on data with engineering interest and because most reliable.
• Use repi because of lack of information on source geometries for many events.
• Only use data from repi 2km because of minimum error of 2km in repi. Exclude data from repi > 100km following arguments of Ambraseys et al. (2005a) (see Section 2.234).
• Mean Mw of the data is 5.39 and mean repi of the data is 45.41km.
• Use vector sum of both horizontal components because geometric or arithmetic means underestimate peak motion.
• Linear baseline correct records. Bandpass filter using acausal 4th-order Butterworth filter with cut-offs chosen based on visual inspection of Fourier amplitude spectra. Low cut-off frequencies between 0.1 and 1Hz, which may be different for the 3 components. Use a uniform high cut-off frequency of 25Hz.
• Use Prefix Gene Expression Programming (using software HSGEP), a form of genetic algorithm, to find the most appropriate functional form based on the independent variables and various mathematical operators. Find models that give highest fitness (lower error, using various measures).
• Use a random 80% in training phase and the remaining 20% in the testing phase.
• Because of a lack of near-source data magnitude-distance saturation not apparent.
• Examine residuals w.r.t. repi and Mw for both training and testing sets and find no trends.