I am not sure if this is the issue, but it is the first thing that jumped at me when I saw your schematic.
You are using large ceramic capacitors in conjunction with paralleled Tantalum or Electrolythics for your bulk supply capacitance, although at first glance this might seem like a good idea, the input characteristics of a switching power supply are very likely to create unstable conditions at their inputs if some precautions are not taken.
The reason can be gleaned from this answer and the exact reason for your specific problem is very probably the exact same one that is shown here.
There is such thing as too little ESR to adequately compensate for the negative impedance presented by a switching power supply. When using ceramics or too low of an ESR for this type of application, you are likely to introduce an undamped resonances that can multiply the input voltage until it reaches destructive levels.
You have to pay attention to the ESR and provide enough dissipation elements (e.g., intentional series resistance) to dampen any oscillation that can be elicited by the circuitry. The combination of your inductor, wiring, and capacitance create a resonant tank that can be excited by the switcher and interacts with its negative impedance. If there is not enough resistance to dampen this oscillation an unstable condition arises which can multiply the input voltage by a large factor.
In your case assuming 50W load with an input of 50V your switcher could present a negative input impedance of -50Ω!!! At 25V it would be -12.5Ω!!! (this of course is a linearized over-simplification, but it illustrates the magnitude of the problem). If I only consider your ceramics, I get a resonant frequency of just 10kHz if I add your electrolytics it goes down to just 1kHz, either of those is below the operating frequency of your switcher which suggests that the linearized approximation would apply!!!
The cause of the negative impedance, your switcher feedback loop, will try to compensate for the oscillating input voltage which will interact non-linearly with the oscillations, very likely making it worse.
You should reconsider the amount of bulk decoupling you are using and test by looking at the input voltage (the lowest you can apply) with enough series resistance in your supply to avoid the whole system from blowing up. If you see large sustained oscillations at the input you would have found the problem.
This can be solved/improved by:
- Adding some series resistance to your bulk capacitance. In most cases less than 1Ω should suffice, but given the large values of negative resistance that might be present, you should analyze the circuit in more detail to be sure.
- Reducing the ceramic capacitors to place the resonance well above the switching frequency, so that it does not interact with your switcher.
- Reducing the overall bulk capacitance, so as the resonance does not interact negatively with your switcher.
- Increasing the wire gauge of the supply, reducing their length, and twisting them to reduce their parasitic inductance.
- Changing the switching frequency. Given the chaotic aspects of the feedback it is likely you would be able to find a stable region of parameter space.
- Modifying the switcher feedback loop, for the same reasons as above.
