update statement of need section

Author: radka-j

paper/paper.bib

@@ -114,4 +114,22 @@ @article{emukit
   author={Andrei Paleyes and Maren Mahsereci and Neil D. Lawrence},
   journal={Proceedings of the Python in Science Conference},
   year={2023}
+}
+
+@inbook{openturns,
+   title={OpenTURNS: An Industrial Software for Uncertainty Quantification in Simulation},
+   ISBN={9783319112596},
+   url={http://dx.doi.org/10.1007/978-3-319-11259-6_64-1},
+   DOI={10.1007/978-3-319-11259-6_64-1},
+   booktitle={Handbook of Uncertainty Quantification},
+   publisher={Springer International Publishing},
+   author={Baudin, Michaël and Dutfoy, Anne and Iooss, Bertrand and Popelin, Anne-Laure},
+   year={2015},
+   pages={1–38} }
+
+   @article{bayesflow,
+  title={{BayesFlow} 2: Multi-backend amortized {B}ayesian inference in Python},
+  author={Kühmichel, Lars and Huang, Jerry M and Pratz, Valentin and Arruda, Jonas and Olischläger, Hans and Habermann, Daniel and Kucharsky, Simon and Elsemüller, Lasse and Mishra, Aayush and Bracher, Niels and Jedhoff, Svenja and Schmitt, Marvin and Bürkner, Paul-Christian and Radev, Stefan T},
+  journal={arXiv preprint arXiv:2602.07098},
+  year={2026}
 }

paper/paper.md

@@ -97,12 +97,9 @@ The latest version of AutoEmulate now also supports direct integration of custom
 
 # State of the field
 
-Compared to other surrogate modelling tools, such as SMT [@smt] or Emukit [@emukit], AutoEmulate [@autoemulate] fills a unique gap in the ecosystem by focusing on making emulation accessible to domain researchers unfamiliar with ML through its automated model selection features.
-
-In this refactor, we have extended the capabilities of AutoEmulate to cover a wider range of emulation workflows, while still primarily targetting a non-ML audience. Autoemulate was originally developed with scikit-learn as the backend. Refactoring the backend to be PyTorch-first allowed us to leverage the wider PyTorch ecosystem as well as the benefits of having end-to-end automatically differentiable emulators and GPU acceleration. AutoEmulate makes use of existing tools that already target parts of the emulation workflow such as SALib [@salib] for sensitivity analysis or Pyro [@pyro] for calibration (we chose Pyro from the available calibration tools for its seamless integration with our PyTorch backend).
-
-The refactor has resulted in a tool that exposes a wide range of emulation capabilities (e.g., surrogate modelling, sensitivity analysis, calibration, active learning) to non-ML experts and translated to a significant growth of the user base and package contributors. We have also retained support for some of the non-PyTorch features following discussions with the community (e.g., the users can still opt in to fit classic ML models such as SVMs or tree based methods although this results in loss of compability with some of the more advanced features).
+The primary motivation for developing AutoEmulate [@autoemulate] was to make emulation accessible to domain researchers who are not machine learning experts through its low code mode and automated model selection features. In this version, the original scikit-learn backend has been refactored to be PyTorch-first, which has enabled us to leverage the wider PyTorch ecosystem and the benefits of GPU acceleration and automatic differentiation. Together, these features distinguish AutoEmulate from other surrogate modelling tools, such as SMT [@smt] or Emukit [@emukit], which do not have the same focus on accessibility for non-ML experts or the same level of integration with the PyTorch ecosystem.
 
+In this refactor, we have additionally extended the capabilities of AutoEmulate to expose a wider range of downstream tasks, while still primarily targetting a non-ML audience. To achieve this, AutoEmulate builds on top of widely adopted tools such as SALib [@salib] for sensitivity analysis or Pyro [@pyro] for calibration as it integrates with the PyTorch backend. Alternative packages exist that also target parts of the emulation workflow. such as SBI [@sbi] or BayesFlow [@bayesflow] which can train emulators on simulated data and perform Bayesian calibration, but they do not cover the same range of emulation capabilities. There are also more comprehensive uncertainty quantification toolkits that offer a range of emulation functionality like OpenTURNS [@openturns] or PyApprox [@pyapprox], however these are not focused on accessibility for non-ML experts. To our knowledge, no existing tool combines automated model selection and hyperparameter tuning with sensitivity analysis, calibration, and active learning in a single low-code package.
 
 # Software Design