This minisymposium seeks to bring together researchers in scientific machine learning (SciML) applied to challenging applications across various domains. In particular, this session will emphasize talks from leading SciML scholars that not only extol the virtues of their developed algorithms but also encourage an expression of their frustrations with the utilization of these algorithms in place of traditional methods of scientific computing. In addition to examples of successful algorithm development, speakers will be encouraged to share incidents of SciML failure, for example, a lack of generalization, limited scalability for realistic applications, model opacity, and the lack of a-posteriori error quantification. Through this, the MS will generate critical discourse on identifying the key algorithmic bottlenecks that need to be addressed for significant advances in utilizing SciML in realistic applications.