Generative artificial intelligence and soil engineering classification: An experiment with Claude (Anthropic)

While traditional artificial intelligence requires to be trained for a given kind of task, requiring (usually large) training data sets for that goal, generative artificial intelligence promises to be a know-it-all, do-it-all, technology, ready to use anytime on any known task (and even tasks never tried before), which brings significative advantages to end users, even those with very limited computer knowledge (essentially, in the best scenarios, the user only needs to launch the program and make requests in natural language). Generative artificial intelligence tools have the potential to centralize and streamline many repetitive tasks with variable complexity, including those requiring chains of diverse logical steps as happens in general with classification procedures. The unified soil classification system for engineering purposes will constitute an interesting case for assessing this potential. In this work we test Claude (from Anthropic) on the diverse tasks required by this classification, using simulated data corresponding to inorganic soils with percentage of fines between 5% and 12%, in order to test for the different types of criteria involved. From discrete (discontinuous) sieve data, the proportions of dimensional fractions (gravel, sand and fines) are computed and grainsize percentiles (D60, D30 and D10) are estimated (Claude was able to perform diverse interpolation techniques and compare the results). From these percentiles are calculated the coefficients of uniformity and curvature. Results of liquid limit and plastic limit are used to calculate the plasticity index, which (with the liquid limit) will be compared to the A-line (as well as the 50% boundary on the liquid limit) to decide whether the fine fraction is silt or clay (with low or high plasticity). The results of all those tasks need to be combined for the soil symbol and name. Additionally, we asked Claude to prepare grainsize plots with marks corresponding to the equivalent diameters (this was the least successful part). Claude was able to treat data from several soil samples on a spreadsheet and prepare a file with the different pertinent parameters and plots. Results were promising for an automatic soil classification by a simple textual request.