Some challenges of calibrating differentiable agent-based models