### 2.381 Derras et al. (2014)

• Ground-motion model is not given here since it requires evaluation of a matrix equation that cannot be summarised. Derive model using feed-forward artifical neural network (1 5-neuron hidden layer; 1 neuron for each independent parameter considered: Mw, log(rjb), log(V s,30), focal depth and mechanism class) with a procedure similar to random-effects approach to compute inter- and intra-event σ. Because provide the matrices and functional form to evaluate model it is included in this section rather than simply being listed. Authors provide spreadsheet to evaluate model. Standard deviations (in terms of common logarithms): τ = 0.155 (inter-event), ϕ = 0.267 (intra-event) and σ = 0.309.
• Use V s,30 to characterise sites. Only 6.9% of sites have V s,30 > 800ms. Italian sites have higher average V s,30 (496ms) than Turkish sites (389ms).
• Use 3 mechanisms (classified using plunge and rake angles):
Normal
540 records.
Reverse
93 records.
Strike-slip
455 records.

Most (76%) Italian events are normal and most (57%) Turkish events are strike-slip.

• Select records from events with focal depth 25km and measured values of V s,30.
• Derive using RESORCE (Akkar et al.2014d) as part of special issue (Douglas2014) including 4 other ground-motion models (Douglas et al.2014).
• Most data from Turkey and Italy. Most Turkish data from rjb > 30km and larger magnitude range than Italian data.
• Data roughly uniformly distributed w.r.t. Mw and rjb for Mw 6. Few larger events.
• Find increasing the number of hidden layers would risk the problem of over-determination without a significant decrease in σ. Find that including depth and mechanism leads to marginal decrease in σ but are included to aid comparisons with other models. Undertake various tests to find most appropriate model.
• Note that despite not imposing a functional form the model is physically sound and suggests nonlinear magnitude, distance and V s,30 scaling.
• Examine inter-event residuals w.r.t. Mw and intra-event residuals w.r.t. rjb and V s,30 grouped by principal country of origin (Italy, Turkey or other). Compute mean residuals by bins and generally find no evidence for bias by country or significant trends. Find bias for Italian intra-event residuals in range 400 V s,30 600ms, which may indicate that V s,30 is not a universal proxy for site amplification (the sites might include those with shallow soft soil overlying hard bedrock).
• Recommend that model is never used outside these ranges of applicability: 4 Mw 7, 5 rjb 200km, 200 V s,30 800ms and 0.01 PGA 10ms. Also recommend model is not used for reverse-faulting events because lack of data.