CorticalG and PhysioG predict EEG and physiological responses to gravity, while Claude 3.5 Sonnet generates first-person narratives of altered-gravity awareness

Introduces and names "gravity-awareness" as an integrated, system-level construct that unites neurophysiological, sensorimotor-behavioral and perceptual-cognitive domains into one computationally modelable response framework. Earlier work treated gravity perception, gravitational orientation and embodied spatial awareness separately; this paper consolidates them into a single operational model intended for computational modelling and predictive assessment. Conceptual synthesis and literature review, illustrated by Figure 1; no independent empirical test of the framework itself is reported.

CorticalG, a Fourier-feature-augmented MLP with two 64-unit hidden layers, maps a single g-load input to percentage change in four EEG bands, achieving test-set MAE of 2.87% for alpha-DMN, 3.15% for beta-PFC, 2.54% for mu-SMN and 1.98% for gamma, with a weighted Huber test loss of 0.2315. Unlike prior work classifying discrete gravity states or comparing single conditions, the model provides smooth interpolation across a continuous 0-1.8 g range. Trained and tested on a synthetic EEG dataset of 94,000 simulated observations from 20 synthetic participants (64 channels, 250 Hz, 2 s windows, 0.5 s stride); the authors report that simulated power distributions and inter-band correlations align with known alpha suppression and beta/gamma enhancement, but empirical flight EEG validation is not yet included.

PhysioG uses eleven independent Gaussian process regressors over cardiovascular, autonomic and motor targets, yielding increased SDNN and RMSSD with reduced heart rate in microgravity, a drop of over 60% in HF SBP in microgravity, a V-shaped skin-conductance response lowest at 1.0 g and rising at both extremes (right SCL over +200% at 1.8 g), and trunk activity peaking at 1.0 g and falling over 50% at 0 g while wrist activity stays relatively stable. Rather than studying single markers, the model delivers multi-system response curves with uncertainty estimates within one framework, validated by leave-one-subject-out cross-validation. Anchor data come from 3-6 discrete g-levels in parabolic-flight literature, interpolated piecewise-linearly into a continuous 0-1.8 g dataset; this is literature synthesis rather than new measurement.

Claude 3.5 Sonnet generated first-person narratives for six gravity scenarios under fixed physiological parameters with only gravitational load varied, producing a coherent gradient: heightened alertness with disorientation in microgravity, greater attentional demand for motor coordination in partial gravity, and narrowed self-awareness under acute strain in extreme hypergravity. The authors describe applying an LLM to simulate embodied awareness under altered gravity as novel, and report that its outputs align in direction with the two quantitative models. A theoretically grounded thought experiment; model choice rested on the authors' subjective comparison of outputs from four LLMs, and the paper states the simulation cannot substitute for empirical data, noting the LLM predicted a ~1.8 m lunar stride without flagging the walk-to-bound gait transition.

The framework targets settings where performance and adaptation must be assessed under altered gravity; the authors envisage real-time EEG monitoring during high-G centrifuge or VR gravity simulations, Gaussian-process tracking of multi-system responses with uncertainty estimates, training difficulty adapted to individual neurophysiological profiles, and personalized pipelines spanning gravity-adaptation schedules, closed-loop VR or centrifuge exposure, neurofeedback and curriculum-based meta-learning. The models are explicitly positioned as quantitative monitoring tools for multi-system signatures of gravity-awareness rather than mechanistic explanations of why alpha rhythms are suppressed in microgravity, and their intended range is continuous interpolation from 0 to 1.8 g, covering microgravity, lunar and Martian partial gravity and nominal hypergravity.

Several open questions remain. The models are trained on literature-derived anchors rather than continuous real-world physiological data, and the authors state that empirical validation with parabolic flight, centrifuge or spaceflight data is needed to assess predictive robustness. The synthetic dataset normalizes each run to its initial 1 g phase and does not capture within-session habituation, in which the first parabola typically elicits stronger responses. Both hypergravity phases were assigned a nominal 1.8 g, whereas the actual pull-out averages about 1.5 g. The models use g-load as the sole independent variable, omitting psychological stress, task demand and individual differences; the data are anchored at g ≤ 1.8, leaving higher-G thresholds that precipitate loss of consciousness and negative effective g-loads unaddressed. The LLM component rests on limited prompts, and the opacity of its training data leaves unclear how far it encodes altered-gravity physiology, while comparability with human self-report awaits validated verbal descriptors. In addition, the loaded text is incomplete: the numeric curves of Figures 4 and 5, the full prompt in Supplementary Section S2 and the complete narratives in Supplementary Section S3 are not included, so statements about intermediate-g behaviour and individual LLM outputs rest on the main-text descriptions alone.