This study investigates the molecular origins of critical-like fluctuations in supercooled water using unsupervised learning. The research focuses on the Liquid-Liquid Critical Point (LLCP) in supercooled water, where two distinct liquid phases coexist. The study uses atomistic simulations and unsupervised learning to analyze the structure of water near the LLCP. The researchers employed local and non-local descriptors to encode information about the molecular environment. Local descriptors, which capture information on a scale of ~3.7 Å, did not reveal evidence for two distinct thermodynamic structures. However, non-local descriptors, which consider fluctuations on a nanometer scale, revealed the emergence of low-density (LD) and high-density (HD) domains, explaining the microscopic origins of density fluctuations near criticality. The study used a three-step protocol involving local atomic descriptors, intrinsic dimension estimation, and free energy landscape analysis. The local SOAP descriptors were used to encode molecular environments, and the intrinsic dimension was estimated to understand the embedding manifold of the data. The free energy landscape was then analyzed to identify minima, revealing that the landscape consists of a single minimum despite macroscopic density fluctuations. This is attributed to the large heterogeneity of local environments in both phases. By expanding the SOAP descriptor to include fluctuations on a length scale of up to 1 nm, the researchers uncovered non-local domains relevant to critical-like fluctuations in supercooled water. The free energy landscape near the critical point evolved between high and low-density macroscopic phases through a complex topography, linked to collective fluctuations of chemical-based order parameters that include non-local information of the water network. The study also examined the relationship between various order parameters and macroscopic density using the Information Imbalance (IB) method. The results showed that descriptors incorporating non-local information, such as the SOAP descriptor and Voronoi density, exhibited strong correlations with macroscopic density. The analysis revealed that the LD and HD domains identified in the study are consistent with the macroscopic density phases, indicating that non-local information is essential for understanding the structural and dynamic properties of water near the LLCP. The findings suggest that the density fluctuations underlying LD-HD transitions cannot be described in terms of local competing structures but involve clusters of at least 100 water molecules. The study provides a framework for understanding water's structural and dynamic properties in scenarios where long-range correlations are important, such as at interfaces and under confinement.This study investigates the molecular origins of critical-like fluctuations in supercooled water using unsupervised learning. The research focuses on the Liquid-Liquid Critical Point (LLCP) in supercooled water, where two distinct liquid phases coexist. The study uses atomistic simulations and unsupervised learning to analyze the structure of water near the LLCP. The researchers employed local and non-local descriptors to encode information about the molecular environment. Local descriptors, which capture information on a scale of ~3.7 Å, did not reveal evidence for two distinct thermodynamic structures. However, non-local descriptors, which consider fluctuations on a nanometer scale, revealed the emergence of low-density (LD) and high-density (HD) domains, explaining the microscopic origins of density fluctuations near criticality. The study used a three-step protocol involving local atomic descriptors, intrinsic dimension estimation, and free energy landscape analysis. The local SOAP descriptors were used to encode molecular environments, and the intrinsic dimension was estimated to understand the embedding manifold of the data. The free energy landscape was then analyzed to identify minima, revealing that the landscape consists of a single minimum despite macroscopic density fluctuations. This is attributed to the large heterogeneity of local environments in both phases. By expanding the SOAP descriptor to include fluctuations on a length scale of up to 1 nm, the researchers uncovered non-local domains relevant to critical-like fluctuations in supercooled water. The free energy landscape near the critical point evolved between high and low-density macroscopic phases through a complex topography, linked to collective fluctuations of chemical-based order parameters that include non-local information of the water network. The study also examined the relationship between various order parameters and macroscopic density using the Information Imbalance (IB) method. The results showed that descriptors incorporating non-local information, such as the SOAP descriptor and Voronoi density, exhibited strong correlations with macroscopic density. The analysis revealed that the LD and HD domains identified in the study are consistent with the macroscopic density phases, indicating that non-local information is essential for understanding the structural and dynamic properties of water near the LLCP. The findings suggest that the density fluctuations underlying LD-HD transitions cannot be described in terms of local competing structures but involve clusters of at least 100 water molecules. The study provides a framework for understanding water's structural and dynamic properties in scenarios where long-range correlations are important, such as at interfaces and under confinement.