THE QUANTUM INTELLIGENCER

A paper deposited on the arXiv describes a saturable absorber formed from the double-transition-metal MXene Mo2TiC2Tx, integrated on a silicon rib waveguide as a nonlinear activation element for photonic neural networks. The authors' Z-scan measurements at 800 nanometres, on films of two thicknesses, yield effective nonlinear absorption coefficients between approximately −2.69×10^3 and −0.88×10^3 centimetres per gigawatt, with the response weakening at higher excitation intensities. The nonlinear absorption is reported as roughly an order of magnitude larger than that of the titanium MXene Ti3C2Tx under comparable conditions. In a neural-network emulator using this activation element, the paper reports 98.39 per cent classification accuracy on the MNIST benchmark. The work is presented as a preprint, and the house will treat it as such: the accuracy figure is a claim awaiting replication, not an established performance. Nor does the abstract establish an operating optical neural network; it establishes a single activation element and an emulator's use of it. The distinction is worth preserving, and the element itself is worth cataloguing. The paper's claims arrive in an order worth preserving. It opens with a materials measurement, proceeds to a comparison against the established titanium composition, then to a waveguide integration, and only at the end to a benchmark figure obtained in an emulator rather than in a physical network. Confidence should taper in the same order. The measurement is the foundation; the comparison is a reason to pay attention; the integration shows that the material can be handled as a compact element; and the emulator result is a suggestion of what might be possible, not a demonstration of a working optical neural network. Prudent engineers will catalogue the material accordingly, and will not yet order the network. What the paper does not establish is worth setting down with the same care as what it does. The nonlinear coefficients were measured at a single wavelength, 800 nanometres, on films of two thicknesses; they do not establish the material's response at telecommunication wavelengths, nor the behaviour of the integrated rib waveguide, for which the abstract reports no equivalent measurement. The comparison with Ti3C2Tx is narrow in the same way: an order of magnitude larger at this wavelength, under comparable conditions, in films. It does not establish general superiority of the double-transition-metal composition. Finally, the MNIST figure describes an emulator. The optical element did not classify digits; a mathematical model containing an activation function of the reported kind did. That distinction is not pedantry. It is the difference between a material that has been measured and a system that has been built. The scope of this result is narrow, and the authors do not pretend otherwise. The abstract reports two film thicknesses, a single excitation wavelength, a single waveguide geometry, and a single benchmark, an emulator rather than a physical network. That is a working envelope, not a demonstration. A material may be measured at 800 nanometres and prove serviceable there without thereby speaking for the telecommunication bands where a photonic network would earn its keep. The comparison with Ti3C2Tx is likewise one comparison at one wavelength; it tells a prudent engineer that the double-transition-metal composition deserves a wider characterisation, not that it has won a contest. To call the work preliminary would imply it aspires to something larger. To call it narrow is only to state what it covers. Within its envelope it is coherent, and the authors have measured what they claim to have measured. Narrowness of this kind is the normal shape of honest materials science: the field advances one wavelength at a time, and a paper that knows its own limits is more useful than one that pretends to none. —Ada H. Pemberley Dispatch from The Prepared E0