(25g) The Effect of Electrochemical Lithium Insertion on the Electronic Conductivity of TiO2 (anatase) and Its Application in Neuromorphic Computing

Artificial neural networks are extremely power intensive because of the need to pass information between digital logic and memory, a result of the von Neumann bottleneck. Analog devices which combine logic and memory functions therefore provide an attractive alternative for neuromorphic computation. Thin-film Li-ion battery materials are attractive candidates for analog transistors due to low switching voltages and a highly linear response. Such devices dynamically insert and remove Li dopants from a host material to modulate the electronic conductivity. However, the electrical properties for many materials as a function of the lithium stoichiometry are poorly understood.

Here, we characterize the electronic conductivity of Li_XTiO₂ (0<X<0.5) as a function of the lithium stoichiometry X upon electrochemical ion insertion. At low lithium stoichiometries (X<0.1), lithium acts as an n-type donor that reduces the Ti⁺⁴ cation and creates a weakly localized electron that increases the electronic conductivity. At higher lithium stoichiometries, however, Li_XTiO₂ transforms from the tetragonal anatase to the orthorhombic titanate phase. This phase transformation counteracts the effect of doping and causes a reduction in the electronic conductivity upon further lithium insertion. Such results are corroborated operando Raman spectroscopy which shows emergence of titanate phase when the conductance drops. We further show that this phase transformation is considerably slower compared to lithium diffusion.

Finally, we demonstrate an artificial synapse based on reversible lithiation of Li_XTiO₂. This microfabricated device operates at very low write voltages, has a linear response, and switches in millisecond time scale. Additionally, this symmetric device does not contain a built-in, open-circuit voltage, and provides one promising avenue for designing next-generation, low-power artificial neural networks.