Please see this post about what to report (e.g. LAMMPS version and platform you are running on, where exactly the run stops) and how to correctly quote input files (otherwise the forum software will interpret special characters as markup, e.g. comments become headlines) and they can become hard to read.
What kind of a “splice correction” are you talking about? I have not heard of that term before.
Some general comment (pun intended): This is an international forum using the English language as a common ground to communicate. Thus it is counterproductive to use comments in your native(?) language when quoting files or providing them as attachments. You are withholding important information from people that do not know your language. English isn’t my native language either, but I learned very early on in my training as a researcher that it is a big advantage to use English for all my notes, comments, and even variable and function names so it would be easier to share and discuss with collaborators.
There are multiple issues here and worrying about the potential function is the least concern. Of course, if you still feel that there could be a problem then you first have to validate that your tables are correct. You would make your life much simpler by using tabulation tool provided in the LAMMPS distribution: 10. Auxiliary tools — LAMMPS documentation. This allows to determine the derivative numerically and thus validate if your force computation for the table and also the format is correct.
Even more so, I would first tabulate the unmodified Buckingham potential and compare it to the output of pair_write.
Some more comments about your input:
Here you are creating completely randomized positions. This is a very bad idea in general. There is no way that the expected borophosphate structure will form correctly. A big problem is that the Buckingham potential is not sufficiently repulsive at short distances to be used for anything but geometries that are already reasonably close to the final structure. A better approach would be to start from templates use molecule files and for cases where different elements show up at equivalent positions, first generate the geometry with just one element and then replace a fraction of them with the other according to the expected stoichiometry and then try to carefully melt and sinter the structure.
What did the people do in the paper that you are referencing to prepare their initial geometries?
Here you are using a timestep of 5 femtoseconds with is quite aggressive for a system with first row elements like boron and oxygen. More reasonable would be 2 fs and conservative (especially since you are starting from high potential energy with your random geometry) would be 1 fs.
Here you are switching to a 1 picosecond timestep. That is completely bonkers and will never work (see above).
In summary, I believe the major problem is your preparation of the initial geometry and your too large timestep is making matters worse. No tweak of the potential function can correct for that.