In recent years, perovskite solar cells (PSCs) have garnered considerable attention as a prime candidate for next-generation photovoltaic technology. Ensuring the structural stability of perovskites is crucial to the operational reliability of these devices. However, the nonphotoactive yellow phase (δ-FAPbI3) of formamidine (FA)-based perovskites is more favorable in thermodynamics, making it challenging to achieve pure α phase in crystallization. Herein, we introduce a language machine learning approach to analyze suitable additives to achieve the desired phase purification. By fast analyzing ∼106 abstracts in chemistry and materials science, our approach identifies aminoguanidine (AG) hydrochloride as a promising candidate for intermediate phase formation during nucleation. The experiments confirm that AG forms a novel one-dimensional intermediate phase (AGPbI3), which suppresses the solvent intermediate phase and δ-phase formation and promotes development of the α-phase. Consequently, the efficiency of the solar cells increased from 23.99 to 25.46%. The long-term thermal stability and photostability were significantly enhanced owing to the purified α-phase, maintaining 82% of the initial efficiency after 1056 h aging at 85 °C and 94.6% of the initial efficiency after 835 h of illumination in an N2 atmosphere, respectively. This strategy also enhanced the performance of flexible solar cells, increasing their efficiency from 21.24 to 22.86%. This work is designed to fast explore new intermediate phases in improving the efficiency and operation stability of thin-film solar cells.