Spooky2 training offers comprehensive customer support and educational materials. No matter your queries regarding hardware, software or the best use for scalars – our knowledgeable instructors have got it covered!
SpookyNet models trained on QM7-X can generalize to much larger chemical structures and excel at geometry optimizations (see methods). Furthermore, it has the ability to accurately reproduce reference potential energy surfaces even when nonlocal interactions are removed from interaction modules.
The architecture
John was inspired to develop Spooky2 by the death of his grandparents and family members, working alongside an international team of electronics engineers, technical designers, software developers and Rife practitioners to produce an advanced and versatile Rife system that would become popular worldwide.
Spooky2 is a flexible modular system that can expand to include additional generators, remotes and coils. It can simultaneously control up to 126 generators independently via its unique theory that connects their lids using a scalar field – acting like an antenna transmitting and receiving antenna of scalar energy.
Spooky2 customers who are new to using our device often become confused regarding the differences between GeneratorX Pro and XM generators, why multiple generators may be necessary for certain diseases, how to access DNA samples into their device, as well as connecting and operating its software. In our training courses specifically designed for you, these topics will be addressed.
Validation
SpookyNet excels on the QM7-X dataset and is capable of performing geometry optimizations on molecules with at least 10 nonhydrogen atoms, even those containing oxygen atoms. In particular, its output produces structures with root mean square deviation from ab initio reference that is either less than or equal to its training error – an impressive accomplishment given that other models evaluated on this dataset include only ML-FFs that do not possess knowledge of chemical properties and depend solely on parameterized functions found in classical FFs.
SpookyNet stands out among ML-FFs by using both atomic numbers and Cartesian coordinates (rather than just energies) as direct inputs for computing its features, similar to many methods that aim to solve the Schrodinger equation54,55,56; it replicates ab initio calculations. Furthermore, SpookyNet encodes angular information explicitly via features based on Bernstein polynomials and spherical harmonics, long utilized in neural network architectures for descriptor creation59-60,61,62,63.
SpookyNet’s energy prediction model includes physically motivated corrections designed to better represent both short-range nuclear repulsion effects as well as long-range electron dispersion effects37,64,65,66,77. These correctives enable it to efficiently provide meaningful chemical insight from single molecules.
SpookyNet can faithfully replicate the reference potential energy surface for all molecules, which is key because models without charge/spin embeddings cannot accurately pinpoint global minima when performing geometry optimizations on unknown structures.
SpookyNet’s prediction performance can be further improved when trained on both fullerenes and lipids datasets simultaneously, as the model becomes exposed to more chemical systems with differing degrees of reactivity. When tested against both datasets simultaneously, SpookyNet was found to produce comparable or better predictions than other ML-FFs on each, producing structures close to its ab initio target structure in terms of relative molecular surface deviation (RMSD; see Figure 2).
