GNGTS 2019 - Atti del 38° Convegno Nazionale

574 GNGTS 2019 S essione 3.1 MODELING OF MAGNETOTELLURIC DATA USING COMPUTATIONAL SWARM INTELLIGENCE: APPLICATIONS TO GEOTHERMAL AREAS F. Pace 1 , A. Santilano 2 , A. Godio 1 , A. Manzella 2 1 Department of Environment, Land and Infrastructure Engineering (DIATI), Politecnico di Torino, Torino, Italy 2 IGG-CNR, Pisa, Italy Introduction. Geothermal areas are usually characterized by a highly complex geology and different kinds of geophysical techniques are adopted to investigate the heat source, the reservoir, its fluid pathways and the impermeable cap-rock. The magnetotelluric (MT) method is one of the most effective geophysical techniques to infer the deep electrical structures of geothermal areas and detect the potential target reservoirs. The inversion of MT data has been historically based on the local search approach, that is, derivative-based inversion methods (e.g., Siripunvaraporn, 2012). However, two main weaknesses of the ill-posed inverse problem have always concerned geophysicists: the final solution 1) can easily find and be trapped in one of the several local minima, and 2) is strongly biased by the initial assumption of the starting model. These two aspects may potentially become of crucial importance for the interpretation of MT data in geothermal areas, in particular cases where external information (e.g., from well- log or other geophysical methods) are unavailable or uncertain. Global-search methods have been introduced in the field of geophysical inversion since they are theoretically able to find, as the final solution, the global minimum of the objective function without being trapped in one of several local minima (Sen and Stoffa, 2013). Computational swarm intelligence (a branch of Artificial Intelligence - AI) adopts the global search approach and includes bio-inspired population-based metaheuristics, such as the algorithm Particle Swarm Optimization (PSO). We implemented the PSO algorithm to perform the two-dimensional (2D) stochastic inverse modeling of MT data (Pace et al. 2019). Our method has proven to be valid to model MT data of very complex geoelectrical structures avoiding a priori initialization. This work presents the first application of 2D optimization (PSO) to MT data coming from the Larderello-Travale geothermal area (Italy). Particle swarm optimization of 2D MT data. The MT inversion requires the minimization of the objective function to find the 2D electrical-resistivity model. The objective function is herein defined to minimize the data misfit between observed data and calculated response and the model roughness. We adopt the optimization with the Occam’s razor principle, that minimizes the effect of spurious features derived by the closest fitting between observed data and calculated response. The high computational cost (in the order of hours) was due to the thousands of MT forward calculations to be performed and was managed by using 24 cores from the High Performance Computing cluster at Politecnico di Torino. The PSO input arguments were accurately calibrated through a sensitive analysis to determine the best values ensuring the solution stability, the convergence speed and an effective minimization of the objective function. For a detailed explanation of the method, readers can refer to Pace et al. (2019). The PSO has been first validated on 2D MT synthetic data sets in order to prove that, differently from traditional methods, the final result of the optimization is independent from the initial assumption of the starting model. Then, PSO was applied to the real-field data set COPROD2, the benchmark test for 2D MT. The optimization of the COPROD2 data set provided a resistivity model of the earth in line with results from previous interpretations, with the striking difference of the random initialization. The MT data set of the Larderello-Travale geothermal area (Italy). The geothermal resource in this area, located in Southern Tuscany, is currently accessed by 18 geothermal plants to produce electric power. The geothermal field is characterized by two different reservoirs (Romagnoli et al. , 2010). The shallow reservoir has been in production for more than a century and is hosted in the mostly electrically conductive carbonate rocks. The deep reservoir, in development since 1980’s and the most used nowadays, comprehends electrically resistive