Fig. 1. a) Experimental setup for BDFA characterization and data sets acquisition; b) Amplifier gain and noise figure as a function of wavelength achieved with 1000 mA pumps currents and −25 dBm signal power; c) Amplifier gain at 1430 nm as a function of total input signal power. TL: tunable laser; MUX: multiplexer; VOA: variable optical attenuator; LD: laser diode; TEC: thermoelectric cooler; Bi: Bi-doped fiber; TFF-WDM: thin film filter wavelength division multiplexer; OSA: optical spectrum analyzer; PM: power meter.

Fig. 2. Neural network architecture for learning the mapping between inputs (signal powers and pump currents) and outputs (gain and NF profiles).

Fig. 3. Probability density functions (PDFs) for gain and NF predictions for a) Case 1; c) Case 2; e) Case 3; the worst and the best gain and NF predictions for b) Case 1; d) Case 2; f) Case 3.

Fig. 4. Maximum absolute error E_MAX of gain and NF predictions as a function of training data set size for three different modeling cases indicated in brackets.