Taking inspiration from the human brain, where memory and learning occur in the same space at a low power consumption, a memristive neuromorphic computing architecture has emerged. However, the integration of low-dimensional materials (2D), such as MXenes, enhances the switching control, spatial and temporal reproducibility, and lower power consumption and fabrication cost. In this work, the fabrication of 2D flakes of titanium carbide (TiC) was performed, through an in-situ HF formation etching. For the synthesis of the TiCflakes 14.78 mL of 9 M hydrochloric acid, 1.6 g of lithium fluoride, and 1 g of TiAlC powder were mixed with 5.82 mL of deionized water. After stirring for 36 h at 35°C, the etched MXene was centrifuged (3500 rpm) in 30 mL of deionized water for 5 min. The washing process was repeated 7 times to achieve the multilayer MXene TiC solution. This suspension was sonicated for 1 h under N, to avoid oxidation. Then it was subsequently centrifuged for 5 min at 3500 rpm to separate delaminated flakes of TiC from multilayer flakes and the supernatant was collected. The MXene was spin-coated on top of a PET/ITO substrate, and 200 µm Ag electrodes sputtered on top.Successful delamination of MXene flakes was confirmed through XRD, SEM, and TEM techniques. Furthermore, the XRD analysis of the dried multilayer and spin coated delaminated MXene films permitted calculation of the interlayer spacing, which is 2.4 Å. This value is important to monitor, because it inherently affects the resistive switching properties of the devices. Furthermore, TEM analysis confirmed the 2D stacking of TiC layers, with -O, -F terminations. The information from the XRD spectra and TEM analysis, allowed us to tune the etching time to 42 h, which gave the most yield but without excess oxidation.The fabricated PET/ITO/TiCT/Ag stack was tested with various current stimulus showing resistive switching at low voltages (<0.3 V). Furthermore, endurance and retention tests showed that the device maintained stability for over 1000 cycles and a ON state for over 30 min (at least). Lastly, long-term potentiation/depression was achieved using pulse stimulation. From these curves, the non-linearity values were extracted, demonstrating a low non-linearity. From the extracted parameters, an hardware online accuracy test was performed with the implementation of a multilayer perception using the Neurosim+ algorithm and MNIST data.
