Discovering new deformation behaviors in multi-material 4D-printed structures is challenging due to the combinatorial size of the design space and the computational cost of high-fidelity solvers. We present a neuro-symbolic discovery framework that couples a graph neural network (GNN) surrogate for thermo-mechanical deformation prediction with a symbolic novelty reasoner that certifies when a candidate response is sufficiently different from previously observed behaviors. The surrogate predicts full-field displacement responses in a few milliseconds per candidate, compared with approximately 30 minutes for a finite element simulation, while achieving MAE = 0.01, MSE = 0.001, and RMSE = 0.03 on held-out evaluation cases. Candidate behaviors are embedded into a descriptor space summarizing deformation characteristics and design features. Novelty is quantified via k-nearest-neighbor distances to reference behaviors; candidates exceeding a novelty threshold are flagged as previously unseen behaviors and clustered into interpretable classes. The symbolic component enforces feasibility constraints and proposes new candidates by targeting under-explored regions of the behavior space, thereby accelerating discovery while keeping novelty claims auditable. Starting from a library of reference behaviors, the framework uncovers multiple clusters of high-novelty responses, illustrating how new material distributions and alignment patterns can induce emergent deformation regimes in 4D-printed structures, an effect widely recognized in programmable liquid crystal elastomers where morphing depends strongly on spatial material/anisotropy programming.