Improve DispBDT reconstruction by adding loss^2 and loss*dist as training variables. This improves sensitivity to truncated images.