I am skeptical that loss for ode works. It is dominant over boundary cost so optimiser tries to go to a local minima where it can zero out all u′′, u’ and u except in the boundary. I had tried it in this question and it seemed following ode cost does not work really even though it has much less λ_ode. The PINN implementation asked several times in this discourse from the searches. You may find your answers in one of them maybe.