We have developed and validated layered analytic models of how high school and university students construct, modify and retain problem solving strategies as they learn to solve science problems online. First, item response theory modeling is used to provide continually refined estimates of problem solving ability as students solve a series of simulations. In parallel, student’s strategies are modeled by self-organizing artificial neural network analysis, using the actions that students take during problem solving as the classifying inputs. This results in strategy maps detailing the qualitative and quantitative differences among problem solving approaches. Hidden Markov Modeling then develops learning trajectories across sequences of performances and results in stochastic models of problem solving progress across sequential strategic stages in the learning process. Using this layered analytical approach we have found that students quickly adopt preferential problem solving strategies, and continue to use them up to four months later. Furthermore, the approach has shown that students working in groups solve a higher percentage of the problems, stabilize their strategic approaches quicker, and use a more limited repertoire of strategies than do students working alone.