A new automatic differentiation approach for fully implicit compositional reservoir simulation

A new stand-alone object-oriented automatic differentiation (AD) tool for FORTRAN 95 codes is presented for facilitating the development of implicit solutions of PDEs. Five different AD approaches were implemented and tested: a forward mode (FM) with static allocation, an FM with dynamic allocation and memory stack, an expression-level reverse mode (RM) with memory stack, an expression-level RM with pointers, and a fully RM with pointers. The new tool is applied to an in-house chemical enhanced oil recovery simulator using three approaches: seed matrix, localised linearisation, and using AD only for computing gridblock properties. The FM AD with static allocation was the fastest AD approach but didn't have the flexibility for problems with variable gradient size. Among the AD coupling techniques, the localised linearisation presented a better performance for assembling the Jacobian when compared to the seed matrix scheme. The use of AD for computing properties only presented the smallest overhead. [Received: May 10, 2023; Accepted: June 1, 2023]