This paper introduces a novel neural network-based computational pipeline for multi-axis 3D printing, which serves as a representation-agnostic slicer. The proposed method enables the generation of curved layers for models with diverse representations and complex topologies. The key idea is to define a mapping function that transforms the input model into a scalar field, which is then used to generate curved layers. The mapping is optimized using loss functions that ensure the layers meet manufacturing objectives such as support-free (SF) and strength reinforcement (SR). The scalar field is derived from the z-component of the mapping, and its gradients are used to define local printing directions (LPDs). The method employs a volumetric mesh as an intermediate representation to compute the mapping, which is independent of the input model's discrete representation. The resulting curved layers are extracted from the isosurfaces of the scalar field and trimmed by the implicit solid of the input model. The method is evaluated on various models, including the Bunny Head, Yoga, and Shelf models, demonstrating its effectiveness in generating curved layers that meet the required manufacturing objectives. The proposed approach is differentiable, allowing for the optimization of the mapping through loss functions directly defined on the scalar field and its gradients. The method is implemented in Python and C++, with PyTorch used for neural network construction and automatic differentiation. The computational pipeline is validated through physical fabrication experiments, showing improved performance compared to existing methods.This paper introduces a novel neural network-based computational pipeline for multi-axis 3D printing, which serves as a representation-agnostic slicer. The proposed method enables the generation of curved layers for models with diverse representations and complex topologies. The key idea is to define a mapping function that transforms the input model into a scalar field, which is then used to generate curved layers. The mapping is optimized using loss functions that ensure the layers meet manufacturing objectives such as support-free (SF) and strength reinforcement (SR). The scalar field is derived from the z-component of the mapping, and its gradients are used to define local printing directions (LPDs). The method employs a volumetric mesh as an intermediate representation to compute the mapping, which is independent of the input model's discrete representation. The resulting curved layers are extracted from the isosurfaces of the scalar field and trimmed by the implicit solid of the input model. The method is evaluated on various models, including the Bunny Head, Yoga, and Shelf models, demonstrating its effectiveness in generating curved layers that meet the required manufacturing objectives. The proposed approach is differentiable, allowing for the optimization of the mapping through loss functions directly defined on the scalar field and its gradients. The method is implemented in Python and C++, with PyTorch used for neural network construction and automatic differentiation. The computational pipeline is validated through physical fabrication experiments, showing improved performance compared to existing methods.