We present the use of interconnected optical mesh networks for early earthquake detection and localization, exploiting the existing terrestrial fiber infrastructure. Employing a waveplate model, we integrate real ground displacement data from seven earthquakes with magnitudes ranging from four to six to simulate the strains within fiber cables and collect a large set of light polarization evolution data. These simulations help to enhance a machine learning model that is trained and validated to detect primary wave arrivals that precede earthquakes' destructive surface waves. The validation results show that the model achieves over 95% accuracy. The machine learning model is then tested against an M4.3 earthquake, exploiting three interconnected mesh networks as a smart sensing grid. Each network is equipped with a sensing fiber placed to correspond with three distinct seismic stations. The objective is to confirm earthquake detection across the interconnected networks, localize the epicenter coordinates via a triangulation method and calculate the fiber-to-epicenter distance. This setup allows early warning generation for municipalities close to the epicenter location, progressing to those further away. The model testing shows a 98% accuracy in detecting primary waves and a one second detection time, affording nearby areas 21 s to take countermeasures, which extends to 57 s in more distant areas. These simulations help to enhance a machine learning model that is trained and validated to detect primary wave arrivals that precede earthquakes’ destructive surface waves. The validation results show that the model achieves over 95% accuracy. The machine learning model is then tested against an M4.3 earthquake, exploiting three interconnected mesh networks as a smart sensing grid. Each network is equipped with a sensing fiber placed to correspond with three distinct seismic stations. The objective is to confirm earthquake detection across the interconnected networks, localize the epicenter coordinates via a triangulation method and calculate the fiber-to-epicenter distance. The ONC will confirm the event and issue early warnings after the third confirmation. According to the Central Italian Apennines (CIA) velocity model [ 34 ], the time window between the primary wave and the arrival of surface wave increases with the increase in the distance from the epicenter, as does the primary wave arrival time. The earthquake struck at 21:41:18 UTC, and the P wave arrived at T0821 after 24 s, MNTV after 28 s and ZCCA after 30 s, as shown in Figure 6 . Consequently, the P wave arrival time is 21:41:42 UTC at T0821, 21:41:46 UTC at MNTV and 21:41:48 UTC at ZCCA. We introduce detailed numbers to show that the time available for early warning in each area is as follows: (We denote the ML P-wave detection time as MLDT and the time difference as TD) In the T0821 area (seconds): (1) Time ( seconds ) = 21 : 42 : 10 − 21 : 41 : 42 + T 0821 MLDT + ( P - wave TD with Z C C A − T 0821 MLDT ) + Z C C A MLDT In the MNTV area (seconds): (2) Time ( seconds ) = 21 : 42 : 24 − 21 : 41 : 46 + M N T V MLDT + ( P - wave TD with Z C C A − M N T V MLDT ) + Z C C A MLDT In the ZCCA area (seconds): (3) Time ( seconds ) = 21 : 42 : 46 − 21 : 41 : 48 + Z C C A MLDT + Zero knowing that Z C C A is the reference station It is good to note that the time difference of primary wave arrivals between T0821 and ZCCA is 6 s (30 − 24), and 2 s (30 − 28) between MNTV and ZCCA. Triangulation Method for Localization Purposes This method is employed by the ONC to pinpoint the earthquake’s epicenter and determine the station/fiber distance from the epicenter to generate early warnings for the nearest area to the epicenter and progress to those further away. The simulator uses measurements of seismic wave arrival times at different stations and their geographical coordinates to estimate the most probable epicenter location by minimizing the differences between expected and observed arrival times. The simulator defines the speed at which the seismic waves propagate through the Earth’s crust, and the coordinates and the exact time at which the wave was detected at each station are specified. The simulator then transforms the wave arrival at each station into seconds relative to the first recorded arrival, and a residual function is employed after to calculate the discrepancies between This is achieved by measuring the distance of each station from a hypothetical epicenter, converting these distances to expected times based on the seismic wave velocity, and then summing the squared differences to create an objective function for optimization. The simulator assumes an initial epicenter positioned at the centroid of the triangle formed by the three stations. A minimization function is then utilized to minimize the calculated residual sum. Table 1 presents a comparative analysis between the actual seismic data recorded by INGV and theoretical detection from the triangulation simulator. The simulator shows almost identical measurements for both the epicenter latitude and longitude. Furthermore, the distances from seismic stations (T0821, MNTV and ZCCA) to the estimated epicenter location exhibit small discrepancies, and this is due to the fact that minor errors are picked up in the peaks from the graphs, and it could also be due to the displacement-to-strain conversion that affects the arrival times of P waves at each station. However, the triangulation method shows efficiency in its estimation. Figure 9 ML detection time of P waves using SOPAS data across three seismic stations/sensing fibers: T0821 ( left ), MNTV ( middle ) and ZCCA ( right ). sensors-24-03041-t001_Table 1 Table 1 Comparison of epicenter locations and distances from seismic stations. Epicenter Location Station to Epicenter Distance (km) Longitude Latitude MNTV ZCCA T0821 INGV Recording 11.251 44.868 47.88 61.45 23.14 Triangulation Simulator 11.2846 44.8705 49.59 63.08 20.48